PDF

Form finding and size optimization
Implementation of beam elements and size optimization in real time

form finding using dynamic relaxation
Masters thesis in Architectural Engineering
JENS OLSSON
Department of Applied Mechanics

Division of Dynamics
CHALMERS UNIVERSITY OF TECHNOLOGY
Gothenburg, Sweden 2012
Masters thesis 2012:46
MASTERS THESIS IN ARCHITECTURAL ENGINEERING

Implementation of beam elements and size optimization in real time form finding using dynamic
relaxation
JENS OLSSON

CHALMERS UNIVERSITY OF TECHNOLOGY

relaxation
JENS OLSSON
JENS OLSSON, 2012
Masters thesis 2012:46

ISSN 1652-8557
Chalmers University of Technology
SE-412 96 Gothenburg
Sweden
Telephone: +46 (0)31-772 1000
Cover:
Shows the results from form-finding and sizing of the roof at Exeter University with SMART Form.
The three overlaid images show the total, bending and axial utilization where red is 70 %.
Chalmers Reproservice
relaxation
Masters thesis in Architectural Engineering
JENS OLSSON
Chalmers University of Technology
ABSTRACT
In the pursuit of performance-based architecture there is an increased demand for integration
between engineers and architects, between analysis and design. The key to make wise choices early
in a design process, has previously been based on experience from prior work and experiments
with physical models. With modern computational power those choices can be supplemented with
real-time feedback from analysis tools. But few programs are designed for the type of light weight,
real-time analysis that can be useful in the exploration of an architectural concept. This has been
the target for a research and development team, SMART Solutions, at Buro Happold.
SMART Form is a tool for performance-based architecture as a plug-in for the 3D modeling software
Rhinoceros. It combines analysis and rationalization of complex geometries with form finding
and sculpting of structures, driven by a dynamic relaxation solver. The plugin transforms the 3D
modeling software to a virtual physical environment, influenced by various forces and a gravitational
field. It enables sketching and sculpting of ecient compression (shells and gridshells) and tension
structures (cable nets and membranes).
The purpose for the thesis has been to advance SMART Form by implementing beam elements and
explore how bending action can be used in the form finding process, allowing for a compromise
between initial geometry and fully form found geometry. Automatic sizing of the beam elements
is also implemented as a method to evaluate form. The thesis leads to an investigation of how to
conduct conceptual structural design, and how this process can potentially be approached dierently
with the means of real-time feedback from analysis, form finding and sizing. In order to allow for
real-time analysis an iterative technique called dynamic relaxation is used as the computational
solver.
As a result of the exploration of form finding with bending action a method called Limit control
form finding was developed. This method enables form optimization of the most critical parts of a
structure by specifying a limit for the bending capability of the elements. It was found to work in a
good way for 2D structures but sometimes resulting in a less appealing, wrinkled looking grids for
3D structures.
Dynamic relaxation is found to be a suitable method for real time analysis, although with some
important issues regarding speed. The number of iterations needed to reach convergence is concluded
to depend on the relation between local and global stiness for each node in the structure. Hence
the method works eciently for some types of structural configuration (for example gridshells with
a locally weak direction), while a conventional matrix solver would perhaps be more suitable for
other cases (for example space truss structures).
The implemented functionality, such as form finding using cables, limit control and automatic
sizing is demonstrated with a couple of test cases. One of which is the Exeter gridshell roof where
engineers at Buro Happold have been involved from the concept design to the realized structure.
The real design process for the project is shown and new approach based on real time feedback
is demonstrated. With the new functionality implemented during the thesis, SMART Form has
become a much more sophisticated tool, and a great improvement for the design process of the
Exeter roof could be made. Due to the possibilities of a quick and ecient collaborative design
process, that comes with the new tool, it seems to oer great potential in bridging between engineers
and architects in structural concept design.
Keywords: Form finding, Dynamic relaxation, Size optimization, Real-time analysis
i
ii
Preface
This Masters Thesis comprises 30 credits and has been carried out during the summer and autumn
of 2012. The work has taken place at the oce for Buro Happold in Bath, England in close
collaboration with the research and development team SMART Solutions.
Main supervisor has been Dr. Al Fisher, analyst and head of research and development in the
SMART Solutions team. From Chalmers University of Technology lecturer Hakan Johansson has
assisted as the examinator for the thesis.
Acknowledgements
First and foremost I want to thank Dr. Al Fisher for his encouragement and guidance throughout
his supervision of this thesis. I would also like to thank Shrikant Sharma, head of SMART Solutions,
for giving me the opportunity to collaborate with Buro Happold for the thesis.
I would also like to thank Dr. Hakan Johansson who has been the examiner from my university for
his support and much appreciated help with the report writing.
Thanks are also due to Prof. Chris Williams, Senior Engineers Mark Cunli and Chris Shrubshall,
whom have all contributed with important feedback to support the software development in the
thesis.
At last I want to thank Erica Henrysson and Marks Stark for the many rewarding discussions we
have had regarding of form finding and coding in SMART Form.
iii
iv
Nomenclature
General note
In general, scalars are written using normal roman letters, e.g. scalar a or scalar A, whilst vectors
and matrices are written using bold roman letters and distinguished between using upper case letters,
respectively, e.g. the vector a and the matrix A. The indices m and i most often refers to element
index and a node index respectively.
Roman upper case letters
A Cross section area

Am Cross section area for element m
C Viscous damping factor
DEA Axial stiness
DEI Rotational stiness
E Youngs modulus
Em Youngs modulus for element m
F1i Global force applied to the node at end 1 (i takes the form x,y,z)
F2i Global force applied to the node at end 2 (i takes the form x,y,z)
G Shear modulus
Hiy Out-of-balance torque for node i around the y-axis
Ix Second moment of area around the local x-axis
Iy Second moment of area around the local y-axis
Jiy Moment of inertia for node i around the local y-axis
Kix Axial stiness for node i in the x-direction
Kv Torsional constant
L Element length
L0 Element initial length
Lt Element length at time t
Lmx Length of element m in the x-direction
Mx1 Moment around local x-axis for end 1
Mx2 Moment around local x-axis for end 2
My1 Moment around local y-axis for end 1
My2 Moment around local y-axis for end 2
M Twisting moment around local z-axis
M1i Global moment applied to the node at end 1 (i takes the form x,y,z)
M2i Global moment applied to the node at end 2 (i takes the form x,y,z)
Mix Mass for node i in the x-direction
t
Mix Mass for node i in the x-direction at time t
Mx,Ed Design values for moment around local x-axis
My,Ed Design values for moment around local y-axis
v
N Axial force
Nm Axial force for element m
NEd Design values for the compression force
Pix External force applied to node i in the x-direction
Pc Buckling force for Euer buckling
Rix Out-of-balance force for node i in the x-direction
S Rotational stiness
Siy Rotational stiness for node i in the y-direction
Tiy External torque applied to node i around the y-axis
U Strain energy
Roman lower case letters
a Acceleration
aix Acceleration for node i in the x-direction
atix Acceleration for node i in the x-direction at time t
c Cross section scale factor
ea Axial deformation due to nodal translations
eb Axial deformation due to nodal rotations
k Eective length scale factor
kxx Interaction factor for axial force and moment
kxy Interaction factor for axial force and moment
kyx Interaction factor for axial force and moment
kyy Interaction factor for axial force and moment
u Parameter for the shape function
ucurrnet Current utilization
usought Sought level of utilization
qy Distributed load applied in the y-direction
v Velocity
vix Velocity for node i in the x-direction
t
vix Velocity for node i in the x-direction at time t
wp Particular solution for the deflection
x Positions along the x-axis in the the global coordinate system
t
x Positions along the x-axis in the the global coordinate system at time t
Roman bold upper case letters
A Transformation matrix for rotations

C Coordinate system written on matrix form
Ji Moment of inertia matrix for node i
K Global stiness matrix
e
K Element stiness matrix
vi
Mi Mass matrix for node i
Roman bold lower case letters
ae Element displacement vector

e
f Element force vector
fl Load vector
r Vector from origin to any point on an element
r1 Vector from origin to the start of an element
r2 Vector from origin to the end of an element
p Vector between the end nodes of an element
y1 Local coordinate axis defining the y-direction for end 1
y2 Local coordinate axis defining the y-direction for end 2
x1 Local coordinate axis defining the x-direction for end 1
x2 Local coordinate axis defining the x-direction for end 2
z1 Local coordinate axis defining the z-direction for end 1
z2 Local coordinate axis defining the z-direction for end 2
Greek lower case letters
Rotational acceleration
iy Rotational acceleration for node i around the local y-axis
t
iy Rotational acceleration for node i around the local y-axis at time t
ix Displacement of node i in the along the x-axis
ijk 3D permutation tensor
Rotation
iy Rotation of node i around the local y-axis
t
iy Rotation of node i around the local y-axis at time t
x1 Rotation of end 1 around the local x-axis
x2 Rotation of end 2 around the local x-axis
y1 Rotation of end 1 around the local y-axis
y2 Rotation of end 2 around the local y-axis
x Curvature around local x-axis
y Curvature around local y-axis
Load scale factor
Non-dimensional slenderness
Material density
Angle of twist
Reduction factor for buckling
LT Reduction factor for lateral torsional buckling
Rotational velocity
iy Rotational velocity for node i around the local y-axis
vii
t
iy Rotational velocity for node i around the local y-axis at time t
ij Rotational velocity for node i in relation to adjacent node j
Greek upper case letters
t Time step
Mx,Ed Moment due to the shift of the centrodial x-axis
My,Ed Moment due to the shift of the centrodial y-axis
viii
Contents
Abstract i
Preface iii
Acknowledgements iii
Nomenclature v
Contents ix
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Free form architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Performance based design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.3 Form finding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.4 SMART Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5 Outline of the report . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2 Theory 14
2.1 Beam theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.1 Coordinate systems and definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.2 Displacements and rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1.3 Local forces and moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1.4 Shape function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.5 Global forces and moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.6 Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 Dynamic relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.1 Translational degrees of freedom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.2 Rotational degrees of freedom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.3 Coupling between rotations and translations . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.4 Algorithm for dynamic relaxation with beam elements . . . . . . . . . . . . . . . . . . 30
2.3 Form finding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4 Beams in form finding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4.1 Property control form finding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4.2 Force/moment control form finding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4.3 Limit control form finding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
ix
2.5 Speed and Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.6 Size optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.6.1 Stress based sizing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.6.2 Virtual work based sizing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.6.3 Stress based utilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.6.4 Buckling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.6.5 Heuristic approach to global buckling . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.6.6 Sizing sections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.7 Load . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.7.1 Distributed load . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.7.2 Surface load . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.8 Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.9 Testing and benchmarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.9.1 Bending test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.9.2 Twisting test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.9.3 Gridshell test case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3 Software 53
3.1 Programming in Rhino . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2 SMART Form User interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4 Results 56
4.1 Simple grid shell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.2 Form improvement using limit control form finding . . . . . . . . . . . . . . . . . . . . . 61
4.3 Exeter University roof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5 Disscusion 77
6 Recommendations for further work 79
References 80
A Appendix A 82
A.1 Derivation of stability condition for rotations . . . . . . . . . . . . . . . . . . . . . . . . 82
x
1 Introduction
1.1 Background
In this chapter topics related to the function of SMART Form as a design tool for performance
driven architecture are introduced. The aim is partly to place the tool in a historical context
among computer tools in architecture and/or engineering and partly to attempt to point out some
limitations with existing tools. Examples of historical and recent projects are presented and this
section is summarized with an introduction to SMART Form.
1.1.1 Free form architecture

In the later half of the 20th century technology took a giant leap forward as computers entered
the scene of science and engineering. Computer aided design allows for visualizing geometries and
analyzing the structural behavior of the most complex systems. Modern 3D CAD software have
become essential tools in any design practice. Whether conducting vehicle, product or architectural
design, the aid of computers have changed not only the work flow but also the design process.
Computer based 3D modeling has historically been influenced from many dierent industries,
although two of the main drivers have been the machine industry, developing cars and airplanes,
and the film and entertainment industry pushing the development of animated illusions. The much
dierent aims with 3D modeling from the two industries led to development of two separate 3D
modeling techniques: surface modeling and polygon modeling. Surface modeling is accurate and
precise whereas polygon modeling is more similar to sculpting with clay, the artistic freedom is
prioritized over precision.
Sir Edmund (Ted) Happold was among the first engineers to ever use computers for modeling in
structural engineering. In the late 1950s his team of engineers at Arup pioneered this new technic in
building design with the Sydney Opera house. The complex design work for the shells was achieved
through the use of computers in order to model the geometry of the roof and analyze its structure
[2]. From the late 50s and until the 90s free form architecture was dominated by form-finding
structures, where there is a distinct relationship between form and function. The development
within this field was led by pioneers like Frei Otto, Heinz Isler, Felix Candela, Ted Happold and
Horst Berger just to mention a few. These structures were probably amongst the most spectacular
at the time (if not still), not only from a structural point of view, but also from a free form point of
view. Some examples of these types of structures can be found in Figure 1.1.2 - Figure 1.1.5.
Computers entered the architectural scene with the PC explosion in the 1990s. To begin with, CAD
was mainly a tool for 2D drawings, essentially just replacing hand drawing, but also introducing a
new level of precision. It was a great step forward in productivity, but did not contribute much to
the design process. Some resisted to adopt the computer modeling, arguing that it challenged the
creativity, whilst other were keen to use this new technic. But it wasnt until 3D CAD was adopted
that computers became a true game changer [7]. The first use of 3D modeling was mainly as a tool
for renderings and presentations, however the significant change occurred when 3D CAD started to
be used as a design tool. The influence of this new way of designing can be seen in projects from the
last decade by architects like Zaha Hadid, Frank Gehry, Studio Morphosis, Asymptote and others
[7]. In comparison to the form finding structures, this computer based sculptural architecture has
seemingly lost the link between form and structural function. With software developed for the film
industry, architectural forms can be sculpted in the most unimaginable ways, tending to not leave
much room for the structural parameters. Two examples of which can be seen in Figure 1.1.6 and
Figure 1.1.7.
Even though computers and 3D CAD have enabled a new type of free form architecture, some
are skeptical to the way these tools are used. In the paper [5] professor Horst Berger raises the
question of what we define as good architecture. Not only the way we conduct design has changed
with the aid of computers, but also what is possible to fabricate and construct, anything seems to
, Applied Mechanics, Masters thesis 2012:46 1

be possible. The almost unlimited availability of tools and technology makes the decision of what
should and should not be designed and built dicult [5]. Berger points out that sensationalism in
itself, often is enough to justify the excessive cost of structurally unreasonable architecture, as a
result of sculptural free form design.
The adoption of CAD software in architecture practices was a slow process but today 2D and 3D
CAD is widely spread in virtually all firms. The next step in the development of CAD systems
is seemingly the exploration of 4th dimension; time [7]. Making time part of the design process,
enables fly-through animations and real time feedback.
One can argue that despite the great availability of computer tools for structural analysis and
architecture, there is still a gap to fill. Especially when it comes to combining architecture and
engineering in conceptual performance based design.
1.1.2 Performance based design
Performance based is a term commonly phrased in contemporary architecture. It implies an urge to

enhance some aspects of performance for a design, a building or more applicable for this thesis, a
structure. The concept of driving design with some aspects of performance is not new, it has been
around for a long time, it is just the label, the tools and the methods that changes.
In order to discuss performance for design it can be helpful to separate and categorize its functions
and features. In the discussion of contemporary architecture the term parameter has become wildly
used. If one thinks of a building as a set of parameters, the design process can be likened to a
balance act, where parameters are weight and prioritized to fit the scheme and context for a project.
Figure 1.1.1: A search at Google Ngram for the expression Performance based. The graph shows the
frequency of use for the expression from year 1900 until 2000.
The parameters dictating the performance of a building are many, intricately linked and impossible
to completely separate. In an attempt, however, to simplify the complexity of building design,
let us consider four categories of parameters: the social, cultural, environmental and economical
category. Some of these can easily be measured and quantified with scientific methods, others require
experience, deep knowledge and practical training to appreciate. The social and cultural aspects of
architecture are categories of parameters that are dicult to measure. The incorporation into the
design process has to be based on intuition, experience and knowledge from previous work. The
environmental and economical category on the other hand, involves quantifiable and measurable
parameters. These aspects are strongly linked to the choice and/or use of material, transportation
and construction technics as well as the use of man-power and technical solutions.
In traditional architectural design the integration of measurable parameters often take place in a late
phase of the design process. The concept design is rather focused on performance of less tangible
nature (aesthetics, experience of spaces and plan functionality), meaning that measurable aspects of
performance often get postponed [25]. Considering the great impact of choices made in conceptual
design this approach is limiting the success of an integrated solution.
In the way the term performance based is used today, there is an almost implicit association with
2 , Applied Mechanics, Masters thesis 2012:46

computer tools together with measurable parameters, and this is no coincides. A computer device
can be programmed to carry out arithmetic and logical operations with a capability of performing
billions of instructions per second. This vast computational power makes them ideal for repetitive
calculations, for instance useful in optimization algorithms. But the target parameters for the
optimization have to be simple enough to be formulated in an computer algorithm. This is typically
the case with measurable and quantifiable parameters. In the application to structural engineering
these could be weight, force, moment, cost or deflection. The parameter in focus for this thesis is
the structural performance, dictated by the use of material for lightweight compression structures.

Historic examples of form finding based architecture
Frei Otto was one of the most influential and recognized architects in the field of lightweight
structures [10]. After he finished his doctoral dissertation in suspended roofs he designed the
Olympic Stadium in Munich in close collaboration with engineers Behnisch and Schlaich. He founded
the Institute of Light Weight Structures in 1964 at the University of Stuttgart and taught the
subject until 1990.
Figure 1.1.2: The Olympic stadium in Munich from 1972. [15]
Felix Candela was trained as an architect at Universidad Politecnica de Madrid, Spain. He got
interested in how calculations could predict strength and collapse of structures. Candela pioneered
innovations in double curved hyperbolic shell surfaces structures, and rather than shaping the
structures arbitrarily, he insisted to design structures where the form corresponds to the flow of
forces [10].
Figure 1.1.3: The Los Manantiales restaurant concrete shell structure from 1956. [22]

Heinz Isler was a Swiss engineer and is often regarded as one of the great pioneers of shell structures.
He gained renown for his experimental form finding methods using physical models. In 1959 at the
first ever IASS (International association for Shell Structures) conference Heinz presented his idea of
shaping shell structures as inverted hanging cloths [10].
Figure 1.1.4: Heimberg Tennis Center, Switzerland
Edmund (Ted) Happold studied Civil Engineering at Leeds University. After graduation he
worked for architects Alvar Aalto and Basil Spence before joining Arup. He became the head of
Structures 3 at Arup in 1967 and worked on landmark projects like the Sydney Opera and Centre
Pompidou. Ted collaborated closely with Frei Otto, Ian Liddell and many other of the pioneers in
lightweight structures. In 1976 when he became a professor in Architecture and Engineering at Bath
university, he left Arup together with 7 other colleagues and started his own firm. The firm is called
Buro Happold and is today a world renown engineering consultancy with over 1400 employees and
27 oces spread out with a global reach [6].
Figure 1.1.5: An interior photograph of Mannheim Multihallen, designed and constructed by Frei
Otto in collaboration with Buro Happold in 1975.

Recent examples of free form architecture
Frank Garyh is a canadian-american architect based in Los Angeles. Garyhs work is said to be
among the most important contemporary architecture and he himself is said to be one of our times
most important architects [26]. Garyh dreamt about becoming a sculptor before choosing the route
of architecture, something that is clearly reflected in his work as an architect.
Figure 1.1.6: The Guggenheim museeum in Bilbao which is among the first examples of computer
based free form architecture [26].
Zaha Hadid is an Iraq-British architect and one of the biggest names within the deconstructionism
style. Already with a degree in mathematics, Zaha moved to London in 1972 to study architecture
at Architectural Association School of Architecture. She joined Oce for Metropolitan Architecture
after her degree and opened her own oce in 1979 [29]. Zaha architects have manifested themselves
as masters of fluid form computer generated architecture.
Figure 1.1.7: The alpine funicular railway designed by Zaha Hadid Architects and opened in 2007
[29]. Seemingly a computer generated sculptural structure, where the form is prioritized over the
structure.

1.1.3 Form finding
The nature of long span, shell and membrane structures relies on the integration of force, function,
geometry and material. The design of this type of light weight structures is heavily constrained
by the laws of nature. In the pursuit of buildings with high-level performance an understanding of
spaces, people and culture, as well as knowledge of how to dictate the flow of forces has to be part
of the early design process.
By minimizing the bending moment and shear force in a structure, less material is needed, leading
to an economical and environmentally sustainable design. The lightness and slenderness of these
structures also give them an aesthetically appealing appearance. However the shaping of such
structure has to undergo a form finding process dictating the resulting shape. Unlike contemporary
free-form architecture, defined purely mathematically (with the aid of computer based 3D CAD),
form-finding shapes are defined by the structures initial geometry, the applied form finding load and
the boundary conditions, making the design process much more involved.
Form finding is a process used to find the geometry for a load-carrying surface stressed structure.
These types of structures are geometrically complex and have to be designed and fabricated with
high precision, in order to work eciently [16]. Form finding has traditionally been carried out using
physical or computer based numerical modeling, often a combination of the two. However computer
based modeling has today taken the over hand since it oers better control and precision. Surface
stressed structures are usually separated into two categories: tensile and compression structures.
Tensile structures referring to membranes or cable nets and compressive structures referring to shells
and grid shells. The form finding process is similar for the two types, however with some important
dierences.
Tensile structures
A tensile structure is a configuration of structural elements that carry load through pure tension.
An example of a tensile structure is membranes, usually constructed as a set of continuous surfaces,
carrying load through membrane action. An other type of tensile structure is cable nets, similar to
membranes in form and function but discretized into cable elements carrying the load through axial
tension. Most tension structures however are combinations of cables and membranes, especially
for larger venues. An example of a tensile structure is the Olympic Stadium in Munich which can
be seen in Figure 1.1.2. Common materials for fabric membranes are PVC, coated polyesters or
PTFE coated fibre glass. Cables on the other hand are usually made out of steel, stainless steel
or some polyester. Tensile structures have the ability to span remarkably long distances with very
little material, making them the most ecient of structures.
Figure 1.1.8: A physical model to form find a membrane geometry using steel wire and soap film
made by Frei Otto. The soap film takes the form of a minimal surface and changes as the boundary
changes [23].
The aim of form finding for a tensile structure is to determine the geometrical configuration where
the loaded membrane does not slack nor crease. This is achieved through an iterative process where
the nodes in the initial geometry are moved in the direction of their out-of-balance force to find
static equilibrium. The force induced on the nodes comes from prestress applied to the adjacent

elements. Membranes are usually constructed with orthotropic fabrics, with dierent properties for
the two main directions (the warp and weft direction). Thus the prestress induced in the membrane
elements can be set to one magnitude for the warp and another one for the weft direction. This
prestress is held constant throughout the form finding process, and since no other loading is applied
the prestress is the governing form driver. If the prestress is set to the same magnitude for the two
directions (which is usually the case), the form found surface will be a true minimal surface.
With the aid of physical modeling, a similar result can be obtained by using soap film or elastic
fabric [16]. By using soap film modeling, the resulting form is close to a minimal surface where the
tension is equal in both directions. With elastic fabrics on the other hand it is the prestress induced
in the two main directions that dictates the form. It can be beneficial with physical models for
designers to experience the stiness and flow of forces. But it is dicult to control the magnitude of
the prestress, making the detailed form elaboration challenging. On top of that, physical models
are dicult to measure and translate into a full scale structure. With the aid of computers this
process can be controlled in more detail, enabling the designer to compromise the form to meet
other constraints. The prestress can be set to dierent values in dierent directions and cables with
a much higher level of stiness can be modeled, enabling a more elaborate design process [16].
Once the form is defined, the structure is analyzed under various load cases by calculating elastic
deformations and stress concentrations in the fabric and the cables. This analysis technic is usually
referred to as elastic control and the technic for form finding by dictating the prestress is called
Force control.

Compressive structures
A shell structure is a double curved continuos solid surface formed to work in pure compression,
usually made from some composite material like concrete. A gridshell on the other hand has
similarities to a shell in form and function but discretized into a grid of bars. These bars can be
made up from any kind of material: steel, aluminum, wood or even cardboard tubes [9]. Generally,
grid shells made from metallic materials are built up with straight members welded at the joints,
whereas grid shells of timber can be made from continuous planks, bent into shape and pinned
together with the crossing planks. Examples of gridshells and a shell structures can be found in
Figure 1.1.5 and Figure 1.1.4, respectively.
The form finding process for compression structures is somewhat dierent compared to form finding
of tensile structures. There is usually no prestress in the elements and the governing driver of the
form is the load (rather then the prestress). The applied load in the form finding process diers
between projects, however, only self weight seems to often be the case. If the structure is a shell or a
grid shell, it is discretized into membrane elements or cable elements. During the form finding process,
the load is inverted (such that it is directed upwards, usually along the positive z-direction), resulting
in a structure that is working in pure tension, since the elements cant take axial compression or
bending. When the equilibrium geometry is obtained, the elements are converted into plates or
beams and the load is inverted (directed downwards, usually along the negative z-direction). For
this particular load case the structure is then working in pure compression and bending moments
and shear forces are minimized. Finally an envelop of load cases accounting for both dead and live
loads are applied, the structure is analyzed and the elements are then sized to cope with the worst
case scenario.
Figure 1.1.9: A reconstruction of physical form finding model made by Antonio Gaudi. The idea is
to distribute the load on the structure as weights hung in chains, the pure tensile hanging geometry
is then inverted forming a compression structure optimized for that particular load case [32].

Structural eciency
Structural performance is an important aspect of free form architecture. A reduction in structural

material often means reduced costs and environmental foot print. However, form finding applied on
free form compression structures tends to strongly dominate the form, sometimes in a way where
other parameters have to be compromised. It is therefore of interest to develop a process to enable
a trade between structural eciency and other architectural design parameters. This is elaborated
in the results section below.
For the case of structures working predominantly in compression, it is a well known fact that axial
force is a more material ecient way of carrying load compared to bending and shear force. Hence to
enhance the eciency of a structure the aim is to reduce the bending moments and shear forces. This
is exactly what form found surface stress structures do, carrying load through axial/membrane action
rather then bending action, making them slender and light weight. But in order to compromise the
form for non-structural parameters, it might be of interest to allow parts of a structure working in
bending whilst other parts are working in axial/membrane action. To be able to make this trade o,
the form finding needs to be carried out using elements that have both bending and axial capability,
i.e. beam elements.
Figure 1.1.10: The initial geometry of an unoptimized structure with two fixed boundaries which are
colored red.
Figure 1.1.11: The geometry for the structure which is form found with the classical hanging chain
model approach. The form finding clearly dominates the resulting form.
Figure 1.1.12: The geometry for the structure which is form found using elements with bending
capability. By controlling the bending capacity it is possible to trade o between initial and fully
form found geometry, allowing for other design parameters to dictate the form.

1.1.4 SMART Form
In the way we conduct the design of architecture or any other product or feature, the the most
important design decisions are the ones made at an early stage [20]. The concept will set the path
for the rest of the design, thus, it can improve the process or make it complicated, which might
result in a lesser and/or more expensive product. Unlike the design industry where the final product
is mass produced and sold in great volumes, each project in architecture tends to be bespoke to fit
a dierent site and context. Whereas product designers can make great revenue by renewing and
optimizing a product on a relative long term basis, an architect or engineer has to invent a unique
building concept and explore its potential at a great speed. Putting the much dierent complexity
of building design into the equation, together with an urge to obtain high level performance, it
stands clear that the two related industries have dierent needs in terms of tools. However many
of the tools used in building design were initially developed for other purposes. The analysis of a
structural concept does not often require the accuracy and complexity of software developed for
cars and airplanes. Likewise, modeling in architecture has other needs of precision compared to
product design or film production. The use of computational power in the architectural design could
potentially be better spent by enabling parametric modeling, combined with real time lightweight
analysis, making it possible to quickly generate and evaluate many permutations of a concept.
The cap in parametric computer tools developed for architecture is something that avantgarde
engineers and architects have worked to fill for the last decade or so. A number of software packages
have been developed for this purpose, including Generative Components and Grasshopper. Today
development is rather focused on integrating analysis into the parametric design environment.
Buro Happold with their research team SMART Solutions are, with their extensive knowledge and
experience in advanced structures, adding on to this development with in-house written analysis
software for performance based design.
SMART Form is a software, written as a free plug-in developed by Buro Happold, for the 3D modeling
software Rhinoceros. It combines analysis and rationalization of complex geometries, with form
finding and sculpting of structures, driven by a dynamic relaxation solver. The plug-in transforms
the 3D modeling software to a virtual physical environment under the influence of various forces and
a gravitational field, enabling sketching and sculpting of ecient compression (gridshells, shells) and
tension structures (cable nets, tents). Prior to this thesis, SMART Form had the capability of doing
conceptual analysis with bar elements. Element properties were calibrated to be scale independent,
rather than controlling material and section sizes the user could control stiness and initial length.
Hence, the software was doing proper structural analysis with bar elements, but did not output any
quantifiable metrics like weight, forces and moments. Loading was controlled by the user setting the
magnitude of the gravitational field.
In essence, SMART Form is a tool for for sculptural design with the foundation in material and
structural principles. The development of the tool takes part of the exploration of 4D CAD, modelling
and analysis based on real time feedback. More information about SMART Form and the latest free
version can be found on [21].

1.2 Purpose
The purpose of this thesis is to advance Buro Happolds real time form finding tool SMART Form,
by integrating beam elements and bending capability into the form finding process, allowing for
a compromise between initial geometry and fully form found geometry. Automatic sizing is then
implemented and the question of how form and size is related is raised and tackled. The thesis
leads to an investigation of how to conduct conceptual structural design and how this process could
potentially be improved or approached dierently, with the means of real time feedback.
In order to conduct structural design in a playful, light weight and conceptual way, real time analysis
is essential. SMART Form is based on dynamic relaxation as a solver, which is beneficial since
intermediate steps between initial geometry and equilibrium geometry are calculated, thus can be
visualized to provide real time feedback. The implementation of beam elements and automatic sizing
will be built around the same dynamic relaxation solver.
1.3 Method
The thesis work started with a general literature review about form finding of structures and the
application of dynamic relaxation to projects within the category of light weight structures. The
focus was to get an understanding of the existing technics for form finding and how they dier
between tensile and compression structure. Gradually, the literature review got more focused on the
mathematics behind dynamic relaxation for bars and beam elements.
The development of SMART Form started with a couple of days of work restructuring the existing code
to enable implementation of new functionality. Thereafter followed a time of software development,
advancing SMART Form by implementing the new types of elements. SMART Form is written as
a plug-in for the 3D modeling software Rhinoceros using Rhino DotNet SDK, based on the C#
programming language. This was the platform for the development. At Buro Happold, Visual
Studio 2005 is used as the programing environment for software development.
During the phase of software development, meetings were arranged with engineers at Buro Happold
to get their feedback into the loop, making sure the tool was developed in a direction that would fit
their design process.
A couple of test cases were developed to test the tool, partly during the development of functionality,
and partly for guidance when designing the user interface. Models from real projects where Buro
Happold have previously been involved, were also used for the same purpose. In order to validate
the implemented elements, a convergence study was performed on three dierent test cases. The
results from SMART Form were then compared with the analytical solutions (for the simplest case)
or with analysis from Matlab (with the finite element toolbox CALFEM) and the commercial FEM
software GSA.
Furthermore elaboration with form finding methods using beam elements was performed and three
dierent approaches implemented. Automatic size optimization was implemented based on knowledge
from discussions with engineers at Buro Happold and additional literature reviews.
Finally the work was consolidated into a finished plug-in with video demos demonstrating the new
functionality in action.
1.4 Limitations
The thesis is focused on solving the statical equilibrium equations for structural analysis using the
iterative technique dynamic relaxation. While geometrical nonlinearity is considered as a result of
the iterative solver, so is not the case for material non-linearity. The implemented material models
are generally simplified to isotropic materials.
The aim with the plug-in is to function as a tool for conceptual structural design rather than

analysis for detailed design. Instead of giving information about the stress distribution in dierent
structural parts, the intent is to give useful feedback on the overall form. Whether a structure is
working in bending or membrane action, roughly what beam sections are needed for a structure
exerted to a certain load, and how the sizes and weight varies when the form of the structure
changes. To be able to give feedback on section sizes, automatic sizing is implemented based on
the methodology for member capacity calculations presented in Eurocode 3. These equations are
simplified to be more suitable for conceptual analysis and partial coecients and safety factor for
loading are neglected. The analysis of beam elements is also limited to closed hollow or solid cross
sections, which furthermore helps to simplify these equations.
Even though programming has been a significant part of the thesis work, it is not presented in detail
in the report, since it is considered as a tool for the development rather than the subject of interest.
1.5 Outline of the report

The first part of the report is an introduction to the topics performance driven design, use of
computers in the design process, and form finding of light weight structures. An attempt is made
to present these subjects from a historical point of view, but also to show their place in a more
contemporary context. The software SMART Form, which is the subject for the development is
then introduced in relation to these topics.
The introduction is followed by a theory chapter, which covers most of the physics and mathematics
used in the structural analysis. The main part consists of an introduction to a vector based beam
theory, showing how to calculate local forces and moments based on displacements and rotations.
It is then shown how these forces and moments are converted to a global coordinate system and
applied to the nodes in a structure. A theoretic foundation is presented for the application of
dynamic relaxation to translational and rotational degrees of freedom. It is then shown how the
static equilibrium can be obtained using dynamic relaxation for beam elements in a high level
algorithm trying to give a holistic view of the process.
The next section in the theory chapter is focused on form finding using beam elements where three
dierent approaches are presented. Finally two inherently dierent approaches to automatic sizing
is touched upon and presented together with theory for calculations of stress based utilization and
local and global buckling. In chapter following, speed and convergence of dynamic relaxation is
treated.
A quick presentation of the user interface for SMART Form is given to bridge between the theory
chapter and the case studies presented in the results chapter.
In the results chapter, it is shown how the new functionality can be used for conceptual structural
design in three dierent case studies. The first one shows how to perform form finding for a simple
grid shell using cable elements and how the height of the roof relates to the need of cross section
sizes i.e. weight of the structure. It is also shown how to quickly evaluate the structural eects of
changing cross sections, materials and orientations. In the second user story it is shown how beam
elements can be used in form finding where a comparison is made between the initial form, a partly
form optimized and a fully form found structure.
The last case study is based on a real project where Buro Happold have been involved in the design
of a grid shell roof. The design process that the engineers and architects went through for the
structural concept, is presented together with a proposal for a new approach based on the new
implemented analysis with real time feedback.
Finally, a discussion about the thesis is given together with recommendations for further work.

2 Theory
This chapter covers most of the physics and mathematics used in the structural analysis. The first
part consists of an introduction to a vector based beam theory, showing how to calculate local forces
and moments based on displacements and rotations. It is then shown how these forces and moments
are converted to a global coordinate system and applied to the nodes in a structure. A theoretic
foundation is presented for the application of dynamic relaxation to translational and rotational
degrees of freedom. Thereafter it is shown how the static equilibrium can be obtained using dynamic
relaxation for beam elements in a high-level algorithm trying to give a holistic view of the process.
Three dierent approaches to form finding using beam elements are presented. And finally, a method
for automatic sizing is implemented and the theory is presented.
2.1 Beam theory

In this section a description of the equations involved in the analysis of beam elements is presented.
A vector based beam theory particularly suitable for 3D dynamic relaxation is introduced. For a
more detailed description of the theory and derivations of the equations presented in this chapter
the reader is referred to Sigrid Adriaenssens PhD thesis Stressed Spline Structures [1].
2.1.1 Coordinate systems and definitions

Each node in the structure is given a coordinate system and 6 degrees of freedom (DOF) by allowing
translations and rotations about the nodal x,y,z axes. Equally each beam is given two separate
coordinate systems, one for each end, giving the element 12 DOF. The orientation of the beam
coordinate systems is measured in reference to a vector p between the ends of the beam. Initially the
z-axis for each of the two beam coordinate systems is aligned with this p-vector. As the coordinate
system of the adjacent nodes rotate, so does the beam end coordinate systems and the resulting
angles, measured in reference to p, causes bending and twisting moments in the element. When
the node coordinate system translates, so does the adjacent beam end coordinate systems, causing
elongation or compression of the members resulting in axial force.
Ym1
zi Z m1
Xm1
yi
Xi
Figure 2.1.1: The coordinate system of node i and the coordinate system of end 1 of adjacent beam
m.

y2
y1
m z2
i
z1
x2
x1
Figure 2.1.2: An example of the arrangement of elements and coordinate systems of the truss modeled
with beam elements. Each element adjacent to node i will contribute with force and moments at the
node.
The forces and moments generated in the elements are applied to the nodes and summed to give
out-of-balance force and out-of-balance moment, which is driving the rotations and translations of
the next iteration, resulting in new rotations and translations. This is repeated iteratively until the
nodes have reached an equilibrium position where loading and element forces and moments balance
each other out. However, the process is started from an initially unstressed geometry, where it is the
sudden application of load that triggers the process by putting the nodes out of equilibrium. This
way of finding the equilibrium geometry can be described in analogy with the more conventional
matrix formulation based on the element force, stiness and displacement equation
f e = K e ae , (2.1.1)
where
K e is the beam element stiness matrix,

ae is the displacement vector containing translations and rotations,
f e is the load vector containing forces and moments.
For the detailed description of eq. (2.1.1) the reader is referred to [14].
The translations and rotations in vector ae are multiplied with the element stiness matrix K e
giving the element force vector f e . By summing f e with the element load vector f l , the new
out-of-balance force and out-of-balance moments are calculated. These forces and moments are then
used in the dynamic relaxation process to move and rotate the nodes giving the new displacement
vector ae for the next iteration. Rather than assembling a global stiness matrix and solving the
system of equations in one go, dynamic relaxation can be seen as an node-wise iterative approach.

2.1.2 Displacements and rotations
Rotation X
The rotations, x1 and x2 , about the local x-axes of the beam element are calculated according to
y1 p y2 p
x1 = , x2 = . (2.1.2)
L L
y2
y1 x2
z2 x2
z1 P
x1
x1
Figure 2.1.3: The rotational displacement of a beam due to the rotation of its end coordinate systems
around the local x-axes. The dashed line shows the centerline of the element and p is the element
vector between the two ends.
Rotation Y
The rotations, y1 and y2 , about the local y-axes of the beam element are calculated according to
x1 p x2 p
y1 = , y2 = . (2.1.3)
L L
y2
y1 z2
y2
x2
z1
y1
x1
Figure 2.1.4: The rotational displacement of a beam due to the rotation of its end coordinate systems
around the local y-axes. The dashed line shows the centerline of the element and p is the element
vector between the two ends.

Rotation Z
The rotation, , about the local z-axes of the beam element is calculated according to
x1 y 2 x2 y 1
= . (2.1.4)
2
y1
y1 z1
x1
x1 P
z1
Figure 2.1.5: The twist of beam m due to the rotation of its end coordinate systems around the local
z-axes. The vector p is the element vector between the two ends.
Axial displacements
The elongation or shortening of a beam element based on the translational displacements of the end
nodes is calculated at time t according to
L2t L20
ea = . (2.1.5)
2L0
Here
Lt = (xt2 xt1 )2 + (y2t y1t )2 + (z2t z1t )2 . (2.1.6)
The total axial deformation of an element has a contribution from the elongation caused by the
element bowing, which is calculated from
L20 [4(x1
2 2
+ y1 2
) 2(x1 x2 + y1 y2 ) + 4(x1 2
+ y1 )]
eb = . (2.1.7)
2L0
y2
z2
x2
dz
y1
z1
P dx dy
x1
Figure 2.1.6: The deformation of a beam due to translations and rotations of its end coordinate
systems. The vector p is the element vector between the two ends.

2.1.3 Local forces and moments
Subsequently, local forces and moments can be calculated form the rotations and displacements
presented in the previous section. The equations are derived based on the relationship between
stress, strain and material properties expressed in terms of energy. Since energy is a scalar quantity
it is a convenient metric for the formulation of the governing equations for elastic bodies in solid
mechanics. In the absence of energy losses due to friction and heat, the external work applied to an
elastic beam causing it to distort, can be expressed in terms of strain energy, which is a form of
potential energy.
The total strain energy in a beam element caused by its three deformation modes (axial deformation,
bending and twisting) can be written as [1]

EA 2 EIx 2 EIy GJ 2
U= (ea + eb ) + dl + 2 dl + , (2.1.8)
2L0 2 L x 2 L y 2L0
where
E is the Youngs modulus,
G is the shear modulus,

Ix , Iy is the second moment of area around x and y-axis respectively,
A is the cross section area,
x , y is the curvature around x and y-axis respectively,

J is the torsional constant,
L0 is the length of the member,
is the twist angle.
Axial force caused by nodal translations and element bowing is calculated by dierentiating eq. (2.1.8)
with respect to e which yeilds
U EA
N= = (ea + eb ). (2.1.9)
ea L0
Equally, the moment equations are obtained by dierentiating eq. (2.1.8) with respect to the angles
x1 and x2 U
Mx1 N L0 4 1 x1 EIx 4 2 x1
= x1
U = + , (2.1.10)
Mx2 x2 30 1 4 x2 L0 2 4 x2
and for the moments around the y-axes,

U
My1 T L0 4 1 y1 EIy 4 2 y1
= y1
U = + . (2.1.11)
My2 y2 30 1 4 y2 L0 2 4 y2
For the twisting moments, the derivative is taken with respect to the twist angle , yeilding
U GJ
M = = (). (2.1.12)
L0

2.1.4 Shape function
The shape of each beam element can be defined by a cubic shape function in terms of a parameter
u. Consider an element m and let vectors r1 and r2 from the origin define the start and end nodes
of the element respectively. Let the parameter at the start of the elements be u = 0.5 and the
parameter at the end be u = 0.5 [1].
y2
z2
r2
u = 0.5 x2
y1 P
z1
r1 u = -0.5
x1
Figure 2.1.7: Beam element defined by a cubic shape function.
Any vector r from the origin to any point on the element can then be described by the function
2 2
r1 + r2 L z1 1 1 1 L z2 1 1 1
r= + pu + p (u3 + u2 ) + p (u3 u2 + ). (2.1.13)
2 z1 p 2 4 8 z2 p 2 4 8
Dierentiating eq. (2.1.13) with respect to u twice yeilds

2 2
d2 r L z1 L z2
= p (6u + 1) + p (6u 1). (2.1.14)
du2 z1 p z2 p
And the curvature around the x-axis
2 2
y 1 Lz1zp1 p (6u + 1) + y 2 Lz2zp2 p (6u 1)
x = , (2.1.15)
L2
x1 (6u + 1) + x2 (6u 1)
x = . (2.1.16)
L
Equally for the curvature around the y-axis
y1 (6u + 1) + y2 (6u 1)
y = . (2.1.17)
L
The moment and distribution due to bending around the x-axis can be calculated based on the
curvature as
M (u)x = (u)x EIx , (2.1.18)
(2.1.19)
Equally for the moment and shear force around the y-axis
M (u)y = (u)y EIy , (2.1.20)
(2.1.21)

2.1.5 Global forces and moments
In the previous sections it has been shown how to calculate the rotations and translations of the
end nodes for a beam element and how to calculate local forces and moments based on these
displacements. However in order to apply local forces and moments to the nodes a transformation
from to global coordinates has to be carried out.
m Z
y1
T
x1 Fy Y
z1
Fx
Fz X
Figure 2.1.8: The local force T at element m transformed into global forces at a node.
The local axial force is transformed to global coordinates using the expression in eq. (2.1.23). The
first term in the equation is basically a decomposition of the force into x, y and z coordinates,
whereas the other terms are the force contribution from the moments.
The force applied to the first end is calculated from
1
F1i = (N pi + Mx1 y1i My1 x1i + Mx2 y2i My2 x2i ). (2.1.22)
L0
And the force applied to the second end is calculated from
1
F2i = (N pi + Mx1 y1i My1 x1i + Mx2 y2i My2 x2i ). (2.1.23)
L0
Here i takes the form x, y and z.
m Z
y1
M1z
M1y M1x
Mx Y
x1 z1
Mz
My
X
Figure 2.1.9: Local moments at element m transformed to global moments at a node.
The moments applied to the first end are calculated from

pk y1j pk x1j (x1j y2k y1j x2k )
M1i = ijk Mx1 My1 + M . (2.1.24)
L0 L0 2

And for the second end the applied moments are calculated based on

pk y2j pk x2j (x1j y2k y1j x2k )
M2i = ijk Mx2 My2 M . (2.1.25)
L0 L0 2
Here
i, j, k takes the form of 1, 2, 3 (representing x,y,z).
ijk is the 3D permutation tensor.
ijk = 1, if i = 1, j = 2, k = 3 or a permutation of these values.
ijk = -1, if i = 3, j = 2, k = 1 or a permutation of these values.
2.1.6 Rotations
Rotation vector
Any rotation can be defined as a scalar for the angle and a vector b as the axis of rotation.
A vector x0 can be transformed from its initial position (x0 , y0 , z0 ) to its current position xt at
(xt , yt , zt ) using a rotation vector b, where the relation between axis of rotation and angle is

|b| = tan ( ). (2.1.26)
2
xt

x0
Figure 2.1.10: The rotation of a vector x0 to xt through rotation vector b.
Net result of two rotations
In the application of rotation vectors to the beam end coordinate systems, the rotation axes are the
global x, y,zaxesandtheanglesarecalculatedf romeq. (2.1.2), eq. (2.1.3), eq. (2.1.4).W henrotatingthenodesaroundtheglo
z or z, y, x). To get round this issue the concept of nonlinear vector rotations is implemented [1].
(a + b + b a)
c= (2.1.27)
(1 a b)
Application in DR
From eq. (2.1.27) the resulting rotation vector from two rotations is calculated. In order to calculate
the rotation of a node in three dimensions the nonlinear vector rotation is used twice.

rx & ry c
rz & c d
Here
r x , r x , r x are rotation vectors for rotation around each of the global coordinate axes x, y, z.
The resulting rotation vector d is then used to calculate a transformation matrix A according to

1 0 0 d1 d1 d1 d2 d1 d3 0 d3 d2
1 cos () sin()
A = cos () 0 1 0 + d2 d1 d2 d2 d2 d3 + d3 0 d1 , (2.1.28)
|d|2 |d|
0 0 1 d3 d1 d3 d2 d3 d1 d2 d1 0
where is obtained from eq. (2.1.26). If the coordinate system axes are defined as

x = x1 x2 x3 , (2.1.29)

y = y1 y 2 y3 , (2.1.30)

z = z1 z2 z3 , (2.1.31)
the whole coordinate system can be written as

x
C = y , (2.1.32)
z
the transformation of C is done by multiplication with matrix A, eectively executing the rotation.
C = AC T , (2.1.33)

x1 x2 x3 a11 a12 a13 x1 y1 z1
y1 y2 y3 = a21 a22 a23 x2 y2 z2 , (2.1.34)
z1 z2 z3 a31 a32 a33 x3 y3 z3
where the left hand side are the transformed coordinate vectors.

2.2 Dynamic relaxation
2.2.1 Translational degrees of freedom

Dynamic relaxation (DR) is a numerical method for solving the dierential equation of motion
by iteratively stepping through time with small increments t. The method can be applied to
dynamic structural analysis as well as static structural analysis, the later being the case in this
thesis. Historically the use of DR as an explicit solution technique for static analysis was introduced
by A.S. Day [3]. The technique had previously been applied to tidal flow computations. Day and
Otter however reformulated the equations by replacing fluid motion and continuity with damped
structural motion and elasticity. The static solution of structural problems can then be seen as the
limiting equilibrium position of heavily damped structural vibrations. DR is especially well suited
for structural analysis of highly non-linear problems, therefore often applied to analysis of tensile
structures.
Figure 2.2.1: A principle sketch of a truss with node i colored in red and its adjacent elements colored
in blue. The out-of-balance force at node i is the summation of the external load Pi and the element
forces.
The basis of the method is to iteratively calculate the movement of all nodes in a structure from their
initial positions, through time steps t, until their equilibrium positions are found and vibrations
have died out due to artificial damping. The movement of each node is caused by an out-of-balance
force calculated as the sum of the forces acting in the elements adjacent to the node and the applied
load. Newtons second law states that the acceleration of a body is directly proportional to the
bodys net force which lay the foundation for dynamic relaxation. From the out-of balance-force and
the nodal mass the acceleration can be obtained, and by stepping through time the velocity and
movement of each node can consequently be calculated. Each node is moved in order to reduce the
out-of-balance force, and when the total force have reached below a certain limit value, equilibrium
is assumed to be reached. A time-stepping method called Euler forward is used throughout the
thesis.
The basis for dynamic relaxation is Newtons second law of motion stating
F = ma. (2.2.1)
This equation can then be rewritten for the displacement of any node i in the x-direction as
Pix Kix ix = Mix aix . (2.2.2)
Where
Pix is the applied force at node i in the x-direction,

Kix is the stiness at node i in the x-direction,

ix is the displacement of node i in the x-direction,
C is the viscous damping constant at node i,
vix is the velocity of node i in the x-direction,
Mix is the mass at node i in the x-direction,
aix is the acceleration of node i in the x-direction.
The relations given in eq. (2.2.2) apply equally for displacements in x, y and z direction. Introducing
the term R, as the out-of-balance force acting on the node, the left hand side in eq. (2.2.2) can be
written
Rix = Pix Kix ix . (2.2.3)
Inserted into eq. (2.2.2) yeilds

Rix = Mix aix . (2.2.4)
Consider the case in figure (2.2.1) where more than one element contributes with force at node i,
eq. (2.2.3) can then be formulated as
n
E m Am
Rix = Pix + (xi xj ). (2.2.5)
m=1
Lmx
Where
Em is the Youngs modulus for element m,
Am is the cross section area for element m,
Lmx is the length of element m in the x-direction,
(xi xj ) is the x-distance between the start and end node of element m.
Inserting eq. (2.2.5) into eq. (2.2.2)
n
E m Am
Pix + (xi xj ) = Mix aix . (2.2.6)
m=1
Lmx
Clearly eq. (2.5.11) is a component in a system of second order dierential equations, here with the
dependent variable xi , where xj is treated as a constant. In order to solve such ODE, integration has
to be carried out twice with sucient initial conditions. For the application in structural analysis
the initial geometry is known, hence the initial positions x is given at time t = 0, and the initial
velocity is known to be v(t = 0) = 0. This is because the analysis starts from an initial equilibrium
position where no loads are applied to the structure.
The integration is carried out using a time stepping method. First the acceleration for node i is
calculated according to Newtons second law of motion.
n
(t+t) Pix + m=1 ELmmxAm
(xti xtj )
aix = t . (2.2.7)
Mix
From the acceleration, the velocity can be obtained by multiplying the acceleration with a time
step t. The viscous damping scalar C is introduced by multiplication with the velocity from the
previous time step
(t+t) t (t+t)
vix = Cvix + aix t. (2.2.8)

Finally, the positions are updated based on the new velocities
(t+t) (t+t)
xi = xti + vix t. (2.2.9)
(t+t)
The equations are nodally decoupled in the sense that the calculation of position xi for node i
only depends on its own residual force and velocity from the previous iteration [3]. From the new
nodal positions the forces in the links are recalculated and summarized to the nodes, giving the
out-of-balance force driving the movement for the next iteration. If the out-of-balance force for all
unconstrained nodes is suciently close to zero, convergence is reached and the static equilibrium
found.
This method can be applied on a system of bars, springs or cable elements where the main dierence
is how the element stiness is calculated. For a cable, however, the force is set to zero if the member
is compressed since it cannot work in compression. In order to incorporate more complex elements
with bending capability, an extra set of degrees of freedom has to be introduced. This is discussed
in the next section.

2.2.2 Rotational degrees of freedom
In this section the implementation of rotations in dynamic relaxation will be presented. Previously
it has been shown how the equilibrium equations for bars, springs or cable elements with 3 DOF
are formulated and solved using dynamic relaxation. Beam elements however have an additional
3 rotational DOF:s for each node giving the capability to work in bending action. By extending
the bar element formulation with the rotational DOF:s a beam element is obtained. To finalize the
formulation, the coupling eect between the translations and the rotations, i.e. forces and moments,
has to be considered. This is treated more in the next section.
Figure 2.2.2: Principle sketch of a truss with node i colored in red and adjacent elements colored in
blue. The out-of-balance moment (the green arrow) at node i is calculated as the sum of the external
torque Ti and the element moments.
Recalling eq. (2.2.1), which is Newtons second law of motion and the foundation of dynamic relaxation
for the translational DOF:s
F = ma. (2.2.10)
This equation can be rewritten to apply for rotations as
T = J. (2.2.11)
Furthermore, for the rotation of any node i around the z-axis, this equation can be expanded to
Tiy Siy iy = Jiy iy . (2.2.12)
Where
Tiy is the torque at node i around the y-axis,

Jiy is the moment of inertia of node i around the y-axis,
iy is the angular displacement at node i around the y-axis,
iy is the angular acceleration of node i around the y-axis,
iy is the angular velocity of node i around the y-axis,
Siy is the rotational stiness of node i around the y-axis.
The relation presented in eq. (2.2.12) applies equally well to rotations around the x and z-axes.
Introducing H for the out-of-balance torque, the left hand side in eq. (2.2.12) can be written
Hiy = Tiy Siy iy . (2.2.13)

Inserting eq. (2.2.13) into eq. (2.2.12) gives
Hiy = Jiy iy . (2.2.14)
Consider the case shown in figure (2.2.2) where more than one element is connected to node i. The
out-of-balance moment is then
n
Em Imy
Hiy = Tiy (4iy + 2jy ) . (2.2.15)
m=1
Lm
The summation regards the elements adjacent to node i, whereas node j refers to the other end of
beam m. For a derivation of this relation the reader is referred to [1].
Where
Em is the Youngs modulus for the material of beam m,
Lm is the initial length of beam m connected to node i,

Imy is the second moment of area of beam m in bending around the y-axis,
n is the number of elements connected to node i,
iy , jy is the rotation in radians around the y-axis for node i and j respectively.
Inserting eq. (2.2.15) into eq. (2.2.14) yeilds
n
Em Imy
Tiy (4iy + 2jy ) = Jiy iy . (2.2.16)
m=1
Lm
In analogy with the case of translational degrees of freedom the integration is carried out using a
numerical time stepping method. First the angular acceleration is calculated from the out-of-balance
moment
n Em Imy t t

Tiy m=1 Lm (4 iy + 2 jy )
(t+t)
iy = . (2.2.17)
Jiy
From the rotational acceleration the rotational velocity can be obtained. Viscous damping is
introduced as a scale factor C (usually set to 0.95) multiplied with the velocity from the previous
iteration,
(t+t) t (t+t)
iy = Ciy + iy t. (2.2.18)
Finally the new rotation is calculated from

(t+t) t (t+t)
iy = iy + iy t. (2.2.19)

2.2.3 Coupling between rotations and translations
In the two previous sections the implementation of translational and rotational degrees of freedom
have been covered without considering the coupling between the two. In order to formulate the
equations of a 12 DOF beam element solved by dynamic relaxation, the coupling eect has to be
considered. In other words, the eect of moments for the translations and the eect of axial force
for the rotations, has to be accounted for.
Recall equations eq. (2.2.5) and eq. (2.2.15) which describe the out-of-balance force and out-of-
balance moments for the translations and the rotations. In order to introduce the coupling eect
these equations have to be reformulated.
The axial force in a member is caused by elongation. Although in the previous formulation
in eq. (2.2.5) elongation only accounts for the translations of the end nodes. However there is
contribution caused by node rotations which makes the element bow. Consider the case with
displacements along the x-axis and rotations around the y-axis, perpendicular to the x-axis. The
axial force is then calculated as
n
E m Am
Rix = Pix + (ea + eb ) . (2.2.20)
m=1
Lmx
Here ea and eb are elongations due to axial deformation and bowing respectively from eq. (2.1.5)
and eq. (2.1.7).
Similarly an axial force in a bent member will introduce moments at its two nodes. This is accounted
for by replacing eq. (2.2.15) with the expression
n
Em Imy Nm L m
Hiy = Tiy (4iz + 2jy ) + (4iy jy ) . (2.2.21)
m=1
Lm 30
The moment contribution from the axial force can be spotted in the second term including the
axial force Nm . Tiz is the applied torque at node i around the z-axis, the sum regards the moment
contribution from each beam element (from 1 to n) connected to node i, and j denotes the index for
the other end of element m.
By comparison with the more conventional matrix formulation for beam elements, the dierent
parts of eq. (2.2.20) can be recognized as elements in the stiness matrix. For the derivation of the
matrix formulation for geometrically non-linear beam elements the reader is referred to [14].
K e = K e0 + K e . (2.2.22)
In the matrix formulation below a 2D beam element with 6 DOF is considered, this is illustrated in
Figure 2.2.3.
u2 u2
u3 u1 u3 u1
Figure 2.2.3: 2D beam elemnt with 6 DOF.

Where K e0 is the beam element stiness matrix and K e is the change in stiness due to the influence
of an axial force Qx .
Written in matrix form

EA
L 0 0 EA
L 0 0
12EI 6EI
0 0 12EI 6EI
L3 L2 L3 L2
6EI 4EI
0 L2 0 6EI
L2
2EI

K0 =
e
EA
L L .
(2.2.23)
L 0 0 EA 0 0
L
0 12EI 6EI 0 12EI
6EI
L3 L2 L3 L2
6EI 2EI
0 L2 L 0 6EI
L2
4EI
L

0 0 0 0 0 0
0 6 1
0 6
5L 1
5L 10 10

0 1 4L
0 1
10 L
30
e 10 30
K = Q x . (2.2.24)
0 0 0 0 0 0
6 1 6

1
0 5L 10 0 10
5L
1 L 1 4L
0 10 30 0 10 30
The blue highlighted terms in matrices K0e and Ke can be identified from eq. (2.2.21) as
n

E m Imy Nm Lm
Hiy = Tiy Lm (4iz + 2jy ) + 30 (4iy jy ) . (2.2.25)
m=1
Written out for each of the four components in the matrix

4Em Imy 4EI
= , (2.2.26)
Lm L
2Em Imy 4EI
= , (2.2.27)
Lm L
4Fm Lm 4Qx L
= , (2.2.28)
30 30
Fm L m Qx L
= . (2.2.29)
30 30
Equally the green highlighted terms in matrix K0e can be identified from eq. (2.2.20) as
n

E m Am
Rix = Pix + Lmx (ea + eb ) . (2.2.30)
m=1
Written out for one of the four components (they are all the same)
E m Am EA
= . (2.2.31)
L L
The matrix components regarding DOF:s 2 and 4 for the beam element in Figure 2.2.3 are not
considered in the comparison since it was found rather complicated, thus outside of the scope for
the thesis.

*&-/(1'+ %-/ #6, +(" /$* 5 1(-, 4(1' !$ + $*$+$,10
'MJJMUGLE DPMK RFC BCQAPGNRGML MD BWL?KGA PCJ?V?RGML DMP RP?LQJ?RGMLQ ?LB PMR?RGMLQ RMECRFCP UGRF
RFC TCARMP @?QCB @C?K RFCMPW FCPC GQ ?L ?RRCKNR RM EGTC ? FMJGQRGA TGCU MD RFC QMJSRGML NPMACQQ @W
NSRRGLE RFC NGCACQ RMECRFCP GLRM MLC ?JEMPGRFK *RCKQ BCQAPG@C RFC QCRRGLE SN MD RFC ?L?JWQGQ
KMBCJ *RCK GQ RFC PCJ?V?RGML JMMN UFGAF GQ ?ARGTC SLRGJ RFC QMJSRGML F?Q AMLTCPECB *RCK ? GQ ?
JMMN ?PMSLB C?AF MD RFC LMBCQ GL RFC QRPSARSPC UFCPC GRCKQ G V BCQAPG@C RFC NPMACBSPC MD FMU
DMPACQ ?LB KMKCLRQ ?PC A?JASJ?RCB A?SQGLE RFC LMBC RM KMTC RMU?PBQ RFCGP COSGJG@PGSK NMQGRGML
$PC?RC ? ACLRCP JGLC ECMKCRPW DMP RFC QRPSARSPC ?R F?LB

$PC?RC LMBCQ ?LB QMPR RFC RMNMJMEW QSAF RF?R C?AF CJCKCLR ILMUQ MD GRQ ?BH?ACLR LMBCQ ?LB
TGAC TCPQ?
*LRPMBSAC @MSLB?PW AMLBGRGMLQ DMP RFC LMBCQ
4CR CJCKCLR NPMNCPRGCQ GC APMQQ QCARGMLQ K?RCPG?J ?LB MPGCLR?RGMLQ
8FGJC LMR AMLTCPECB
? 'MP C?AF LMBC
G $?JASJ?RC K?QQ ?LB KMKCLR MD GLCPRG? DMP RFC %0' ?AAMPBGLE RM CO ?LB
CO
GG 4CR LMBC DMPACQ ?LB KMKCLRQ RM XCPM
GGG $?JASJ?RC DMPAC ?LB KMKCLRQ @?QCB ML RFC BGQNJ?ACKCLRQ ?LB PMR?RGMLQ DPMK NPCTGMSQ
GRCP?RGML @W JMMNGLE ?PMSLB C?AF @C?K ?BH?ACLR RM RFC ASPPCLR LMBC
$?JASJ?RC RFC KCK@CP CLB PMR?RGMLQ DPMK CO CO ?LB CO
$?JASJ?RC RFC CJMLE?RGML BSC RM ?VG?J BGQNJ?ACKCLR ?LB @MUGLE DPMK CO
?LB CO
#?QCB ML RFC LCU BGQNJ?ACKCLRQ A?JASJ?RC RFC ?VG?J DMPAC DPMK CO
&OS?JJW DPMK RFC PMR?RGMLQ A?JASJ?RC RFC @C?K CLB KMKCLRQ QCC CO ?LB
CO
5P?LQDMPK RFC ?VG?J DMPAC ?LB RFC KMKCLRQ RM EJM@?J AMMPBGL?RCQ ?LB ?NNJW RM
RFC LMBC ?AAMPBGLE RM CO ?LB CO
GT "NNJW JM?BGLE QSAF ?Q QCJD UCGEFR ?LB QSNCPGKNMQCB JM?B A?QCQ ML RM RFC LMBC
T .MTC RFC LMBC GL RFC BGPCARGML MD RFC MSRMD@?J?LAC DMPAC
$?JASJ?RC RFC LMBC ?AACJCP?RGML DPMK CO
$?JASJ?RC RFC LMBC TCJMAGRW DPMK CO
6NB?RC RFC LCU LMBC NMQGRGML SQGLE CO
TG $?JASJ?RC RFC LMBC PMR?RGML ?LEJC @?QCB ML RFC MSRMD@?J?LAC KMKCLR
$?JASJ?RC RFC ?LESJ?P ?AACJCP?RGML MD RFC LMBC ?AAMPBGLE RM CO
$?JASJ?RC RFC ?LESJ?P TCJMAGRW MD RFC LMBC ?AAMPBGLE RM CO
$?JASJ?RC RFC PMR?RGML ?LEJC ?AAMPBGLE RM CO
TGG #?QCB ML RFC EJM@?J ? @ ?LB A ?VCQ ?LB RFC M@R?GLCB PMR?RGML ?LEJCQ A?JASJ?RC RFC
LMLJGLC?P PMR?RGML TCARMP DPMK CO
TGGG $?JASJ?RC RFC RP?LQDMPK?RGML K?RPGV DPMK CO
GV 3MR?RC RFC LMBC @W RP?LQDMPKGLE GRQ AMMPBGL?RC QWQRCK
$?JASJ?RC RFC PCQSJR MD RFC RFPCC PMR?RGMLQ SQGLE CO
$?JASJ?RC RFC RP?LQDMPK?RGML K?RPGV @?QCB ML CO
&VCASRC RFC PMR?RGML @W RP?LQDMPKGLE RFC AMMPBGL?RC QWQRCK DPMK CO
V &OS?JJW RP?LQDMPK RFC CLB AMMPBGL?RC QWQRCKQ DMP RFC ?BH?ACLR @C?K CLBQ
@ 4SKK?PGXC RFC MSRMD@?J?LAC DMPACQ DMP ?JJ RFC SLAMLQRP?GLCB LMBCQ ?LB AMKN?PC UGRF
RFC AMLTCPECLAC APGRCPG? *D RFC DMPAC GQ QK?JJCP RF?L RFC AMLTCPECLAC JGKGR RFC ?L?JWQGQ GQ
BMLC ?LB RFC UFGJC JMMN AMLBGRGML GQ QCR RM D?JQC

*+)* +"*#*
i
Figure 2.2.4: Principle sketch for loop logic presented in the algorithm above. The blue arrow
represents the loop around each of the nodes in item (a), and the green arrow represents the loop
around each element adjacent to node i described in (iii) in the algorithm.
2.3 Form finding
There are essentially two inherently dierent approaches to form finding of structures; elastic control
and force control. Both methods aim to find the equilibrium geometry for a structure under a
certain load and/or prestress, but the way the forces are calculated in the elements diers. A tensile
structure is usually form found with the pretension in the membrane or cables as the governing
form driver and the eect of self weight and loading is neglected. In the analysis of the form found
structure when loading is accounted for the prestress is still the main source of stress in the structure,
hence chosen as the driver in the form finding process. For a compression structure on the other
hand, loading rather than prestress, is the main source of stress in the structure and is used as the
governing form driver.
Form finding of compression structures
Elastic control form finding is basically non-linear structural analysis with the purpose of finding
the static equilibrium geometry for a structure under a certain load. The classic approach for a
compressive structure would be to convert all the bar and/or beam elements to cables and invert the
load (see Figure 2.3.1) such that all elements work in pure tension. The initial geometry together
with the section and material properties of the cables will dictate the resulting equilibrium geometry.
This process can be formulated as a system of equilibrium equations set up and solved using a
matrix formulation. However, the analysis with cable elements is heavily non-linear, meaning that
the matrix solution has to be carried out iteratively. Another approach to find equilibrium is to use
dynamic relaxation which by its iterative nature inherently accounts for geometric non-linearity.
This is the method adopted in this thesis. Dynamic relaxation is based on the dierential equation
for damped motion which describes a relation between acceleration, velocity and distance. The
sudden application of load on the initially resting nodes put them out of equilibrium. Using dynamic
relaxation the acceleration is calculated based on Newtons second law in eq. (2.2.1) where f is
the out-of-balance force. By stepping through time the velocity and distance to move the nodes is
then obtained. As the nodes move, the bars connected to the nodes start to stretch introducing
forces pushing and pulling at the nodes. Each node is moved in the direction of its out-of-balance
force in an iterative process until the applied force from the loading is compensated by bar forces,
meaning that the node has found equilibrium. The mathematics behind the process is presented
more thoroughly in chapter 2.2 Dynamic relaxation.

Figure 2.3.1: Reversed loading (the blue arrows) applied during form finding of a grid shell structure.
The elements are set to cables and a pure tension structure is obtained.
Form finding of tensile structures
Force control form finding is used for tensile structures and uses prestress rather than loading as
the governing form driver. The prestress in the membrane and/or cable elements is preset and
these forces are summed on the nodes giving the out-of-balance force. The nodes are moved in the
direction of the out-of-balance force using the equation of motion solved with dynamic relaxation.
As the nodes move, the bars connected to the nodes start to stretch, but instead of calculating the
force in the bars based on deformations, the force is preset and kept constant throughout the whole
process. For a tensile structure the only driver of the form is this prestress, no external loads are
applied. Fabric membranes usually have dierent stiness in the warp and weft direction based
on the orthotropic properties of the weave. Therefore the prestress can be set dierently for the
two directions and the relationship between these two prestresses is what dictates the resulting
equilibrium geometry. In practice this means that the resulting form is independent of the magnitude
of the prestress but rather dependent on the ratio between the prestress in the two directions. The
process is also independent of material and section properties for the elements.
warp
weft
Figure 2.3.2: Prestress in the weft and warp direction pulling at the nodes. It is the ratio between
these stresses that dictates the result of the force control form finding.

2.4 Beams in form finding
In this thesis form finding with bar, cable and beam elements has been implemented partly based
on the traditional, pure tension hanging chain model approach, and partly in a more experimental
way, influenced by the force control technic used for tensile structures. The aim has been to develop
a method to conduct form finding which allows for a trade-o between initial geometry and fully
form found geometry using elements with bending capability.
Three dierent approaches to form finding with beams are presented below, each of which is based on
a reversed load case as the governing form driver where the aim is to find the equilibrium geometry
of the structure.
2.4.1 Property control form finding

Property control form finding is implemented as a method of controlling axial and bending stiness
for the elements in a structure during the form finding process. Scale factors for axial (EA) and
bending (EI) stiness is exposed in the user interface as slider bars, allowing for the user to elaborate
with the form. By dropping the bending stiness to zero the beams are eectively converted into
bar elements and the form adapts to find equilibrium without bending. By increasing the bending
stiness again the form adapts to work partly in bending and partly in membrane action where this
ratio is controlled by the user. The dierent form options can be saved and evaluated using automatic
sizing. With this approach each element in the structure is eected equally by the property scaling,
whether or not it is heavily utilized and regardless if it is working in axial or bending or action.
2.4.2 Force/moment control form finding

This type of form finding is inspired by the method used for tensile structures. Instead of calculating
the axial force and the moments based on translations and rotations, these values are set based
on the capacity of the elements. The user controls the force in the elements with a slider bar by
specifying the value as a percentage of the axial capacity. The bending moments are controlled in a
similar fashion, based on the percentage of the bending capacity. In contrast to force control used
for tensile structures the governing form driver is a load case. Thus it becomes a tricky balance act
to ensure that the forces set in the elements can balance the applied loading. If the load is large
compared to the element force, equilibrium might never be found. This is because the force does
not change with the elongation of the members, it is just the direction of the force that changes
when the geometry changes.
+ =
Figure 2.4.1: Load (green) applied on a node and the specified forces (blue) in the adjacent elements.
Even if the geometric configuration of the elements changes, equilibrium might not be possible for
this relation between load and force.
If was found a bit unintuitive to set both the x and y-moments of an elements to the same value,
preferably these would be controlled separately. But because of the diculty to balance the forces
this method was not found to be particularly useful. However one great benefit with this method is
that it allows the elements to more easily change from their initial length, since the change of length
does not aect the axial force. This was found to be a problem with other methods, causing the

structure to wrinkle in an undesirable way. The investigation however led to the implementation of
limit control form finding presented in the next section.
2.4.3 Limit control form finding

Limit control form finding has similarities to the two methods presented above. It allows for
elaboration with the ratio of bending and membrane action like in property control, but rather
than scaling the EA and EI values for the elements the user specifies limit values for the axial and
bending utilization in a way more similar to setting the values in force/moment control. These two
limits are percentages of the member capacity, meaning that two beam elements with dierent cross
sections but the same bending utilization limit are allowed to reach dierent maximum moments. If
a member is loaded such that the axial force or the moment exceeds the specified limits the exceeded
quantities are scaled down and the form will have to adapt. This approach makes it possible to
enhance the structural performance in the most critical parts of a structure, where for example the
bending action is dominating, without necessarily changing the parts that are not as critical.
To demonstrate how this works, consider the arc like structure in Figure 2.4.2. The load is reversed
to act as a form finding load and the elements have circular hollow sections with a radius of 0.2 m.
48 m
Figure 2.4.2: Boundary condition, loading and initial center line geometry for a bent arc like
structure.
First of all elastic analysis is carried out showing that the mid part of the structure is far over
utilized (see Figure 2.4.4). Secondly limit control form finding is carried out with the axial capacity
limit kept constant at 50 % and the bending capacity varied from 75% - 0% resulting in 4 dierent
geometries.
% Utilization
100
0 % Bending
25 % Bending
50 % Bending
50
75 % Bending
Elastic analysis
Figure 2.4.3: Equilibrium geometry as a result of elastic analysis and 4 cases of limit control form
finding with various settings. The colors display the utilization of the members where red is 100 %
and blue is 0 %

% Utilization
100
3.48 3.24
3.61
26.58
27.18 3.0
27.07
45.67 18.19
47.09 51.59 26.75 50

56.82
108.54 75.37 52.46

76.7
144.9
115.79 0
144.99
Figure 2.4.4: Utilization values for the critical part of the structure.
The initial structure in the example is due to its form predominantly working in bending which
becomes clear by studying the values in the utilization plot in figure (2.4.4). The most critically
stressed element in the case with 75 % as bending limit is utilizing 76.7 % of its total capacity,
implying a 1.7 % utilization in axial action. Since the axial utilization is not significant in the
example the limit is kept constant at 50 %. When the bending limit is dropped to 0 % the structure
takes the form of a perfect cantenary and by observing the utilization values, now ranging from
about 3.0 - 5.2 %, a remarkable dierence structural eciency is clear. On the other hand the form
has undergone a drastic change under the form finding process potentially hindering other design
parameters.
High level algorithm for the limit control form finding:
1. The dynamic relaxation process runs one iteration, calculating force F and moments M in the
elements.
2. The axial and bending utilization UF and UM are calculated respectively.
3. The utilization limits are sent across from the UI i.e. If the limits are 40% axial and 60%
bending, then UF40% , and UM60%
are calculated.
4. If |UF | > UF40% , then set the force such that |UF | = UF40% is obtained.
60%
5. A similar check is made for the moments, if |UM | > UM , then the moments Mx and My
(giving UM ), for bending around the x, and y axis respectively, are both scaled equally such
60%
that |UM | = UM .
6. The node positions and rotations are now updated according to the calculated and scaled
forces and moments.
Based on the levels of utilization the form adapts to find its equilibrium position for the applied
load. This process of form finding allows for an interesting trade o between bending and membrane
action.

2.5 Speed and Convergence
To achieve the objective of making SMART Form a real time and interactive analysis tool, speed
is essential. The aim is to reduce the time it takes to reach convergence while making sure the
analysis is stable. The obvious approach to increase speed of a dynamic relaxation solution would
be to increase the time step. A greater time step means that each node is moved a greater distance
for each iteration, resulting in lesser iterations needed to reach the equilibrium position, improving
speed, assuming that instability does not occur. Two other parameters that dictate speed is the
mass and moment of inertia used in eq. (2.2.7) and eq. (2.2.17). Lets first consider the mass and
recall the use of Newtons second law which states
F = ma. (2.5.1)
Which is used in dynamic relaxation to calculate the acceleration from eq. (2.2.7), in a simplified
matter this can be written
F
a= . (2.5.2)
m
For a certain out-of-balance force the magnitude of the acceleration is completely dictated by the
choice of the mass, a smaller mass results in a greater acceleration. If the aim is to trace the real
dynamic behavior of the structure, the masses would be lumped based on the real structural weight
[3]. In the case of static analysis however, the interest is solely to reach convergence as quick as
possible, and fictitious masses calibrated for this purpose can be used. In the paper [3] by Michael
Barnes, a derivation is presented for the relation between time step, mass and stiness for a node in
a cable net structure. Barnes proposes to keep the time step as a global constant and calibrating
individual masses for the nodes based on their local stiness. This approach is applied in the thesis,
but rather than calculating one mass per node, one mass for each degree of freedom is calculated
and formulated here as a mass matrix.
The relation between time step, mass and stiness that will ensure a stable solution is formulated as
[3],

2m
t = . (2.5.3)
S
For a constant t the mass is calculated as
t2 S
m= . (2.5.4)
2
Where the mass matrix M i of node i can be expressed as a scalar multiplied with the node stiness
matrix
t2
Mi = Si. (2.5.5)
2
Written on matrix form

Sx
2
0 0
t
Mi = 0 Sy 0 . (2.5.6)
2
0 0 Sz
The components in the stiness matrix S are calculated from

n
E m Am
Sj = |pm cj | . (2.5.7)
m=1
Lm
Here j takes the form x, y, z, the summation over m regards the elements 1 to n adjacent to node
i, c is the unit vector taking the form of the global coordinate axes x, y, z and pm is the vector
between the end nodes of element m.

Equally a relation between time step, moment of inertia and rotational stiness that ensures speed
and stability is needed for the rotational degrees of freedom in eq. (2.2.17). In Appendix A an
attempt to derive that relationship based on Barnes approach for the translational degrees of freedom
is presented. Let us consider the relation between time step, moment of inertia and rotational
stiness as

2J
t = . (2.5.8)
S
Where
S is the rotational stiness,

J is the moment of inertia,
t is the time step.
In analogy with the masses eq. (A.1.14) is used to compute the components of the moment of inertia
matrix J i for node i as
t2
Ji = Si. (2.5.9)
2
On matrix form

2
Sx 0 0
t
Ji = 0 Sy 0 . (2.5.10)
2
0 0 Sz
Where the components in the stiness matrix S are calculated from

n
1
Sj = (Em Imx X j + Em Imy Y j + Gm Kvm Z j ) . (2.5.11)
m=1
L
Here j takes the form x, y, z; the summation over m regards the elements 1 to n adjacent to node i;
X j , Y j , Z j are the normalized local coordinate axes and their component in j-direction; Ix , Iy are
the second moment of area around local x, y axes respectively and Kv is the torsional stiness.

Convergence criterion
The result form numerical calculations applied to a scientific problem will usually be an approximation.
In order to formulate the convergence criterion it is therefor of interest to find a value where the
approximation gives an error that is acceptably small. The level of acceptance may vary depending
on the purpose of the analysis. In the implementation of convergence criteria in SMART Form,
one criterion is used for the sizing whilst another one is used for the general analysis. To improve
the speed of the automatic sizing, the sizing criterion is allowed to be more easily reached. This is
because the static solution has to converge before the sizing is executed, and around 5-10 sizing
iterations has to be performed for convergence.
Time stepping
As shown in the theory in section 2.2 Dynamic relaxation the time stepping algorithm implemented
for the analysis is the Euler Forward method. This is a general method that can be applied to solve
ordinary dierential equations using the formula
yn+1 = yn + hf (xn , yn ) (2.5.12)
The method uses information from the beginning of the interval only, to increment the solution. It
is not considered as being very accurate nor very stable in comparison to more involved methods,
but has the advantage of being simple to implement [30].

2.6 Size optimization
The theory of two dierent approaches to automatic member cross section sizing is introduced in
this chapter. The choice of method depends on what parameters are being optimized for. From a
structural point of view there are two limit states that has to be satisfied in order for a structure
to be classified as safe; the ultimate limit state and the serviceability state. A structure should be
designed such that the probability of reaching one of these limit states, at which it would become
unfit for its intended use, is acceptably low [18]. A stress based approach is used when sizing for
ultimate limit state, essentially making sure that no members in a structure are over/under-utilized
due to stress. Sizing for serviceability state on the other hand is about minimizing deflection at a
critical point in the structure, and a virtual work based method is more suitable in this case. The
stress based sizing is implemented during the thesis, while virtual work based sizing was found less
useful for conceptual analysis.
2.6.1 Stress based sizing

The ultimate limit state concerns strength and stability requirements of a structure and can be
seen as a collapse criterion. None of the members in the structure are allowed to be utilized over
its capacity causing the risk of a collapse. To perform sizing for ultimate limit state, each member
utilization is checked and the cross sections are scaled such that the sought utilization is reached
[18].
A high level algorithm used for stress based size optimization.
1. While (any element is over/under-utilized, i.e. sizing not completed)

Running the analysis:
(a) While (not converged, i.e. analysis not completed)
(b) Run the dynamic relaxation one iteration.
(c) Check the out-of-balance force for convergence.
2. Calculate the utilization of the elements due to axial force and bending moments.
3. Scale the beam to reach sought utilization.
4. Update section based properties like mass and stiness.
2.6.2 Virtual work based sizing

The serviceability limit state concerns a deflection requirement for a structure. If a structure fails to
meet this criterion it might still be strong enough to stand up but it will be unfit for its intended use,
for example in terms of durability or fire resistance. Automatic sizing, to fulfill a deflection criterion,
starts by identifying a critical degree of freedom (DOF) for deflection. A unit load is applied in
that DOF and a virtual work based method is used to calculate which elements that contributes to
stiness. These elements are sized until the deflection criterion is met.
Below is a high level algorithm used for virtual work based size optimization. The main dierences
to the stress based sizing is the convergence criterion and the approach used to find which elements
to size.
1. While (deflection criteria is not fulfilled, i.e. sizing not completed)

Running the analysis:
(a) While (not converged, i.e. analysis not completed)
(b) Run the dynamic relaxation one iteration.

(c) Check the out-of-balance force for convergence.
2. Apply a unit load in the degree of freedom for the critical deflection.
3. Find each elements contribution to stiness in the critical DOF.
4. Scale the element sections in relation to their stiness contribution.
5. Update section based properties like mass and stiness.
2.6.3 Stress based utilization

The stress based utilization of a member is calculated according to Eurocode 3 specifications for
uniform members in bending and axial compression (Ch. 6.3.3). It states that members subjected
to bending and axial compression should satisfy [13]
NEd Mx,Ed + Mx,Ed My,Ed + My,Ed

+ kxx + kxy 1, (2.6.1)
x N Rk
M 1
Mx,Rk
LT
My,Rk

M1 M1
NEd Mx,Ed + Mx,Ed My,Ed + My,Ed

NRk
+ kyx Mx,Rk
+ kyy My,Rk
1. (2.6.2)
y M 1 LT
M1 M1
Here
NEd , Mx,Ed , My,Ed are the design values for the compression force and maximum moments about
the x-x and y-y along the member respectively.
Mx,Ed , My,Ed are the moment due to the shift of the centrodial axis according to [13] (Ch.
6.2.9.3).
x , y are the reduction factors due to flexural buckling [13] (Ch. 6.3.1).
LT is the reduction factor due to lateral torsional buckling [13] (Ch. 6.3.2).
kxx , kxy , kyx , kyy are the interaction factors.
In the implementation in SMART Form the two equations above are simplified, partly to not over
complicate the otherwise lightweight and conceptual analysis and partly to enable a separation of
axial and bending utilization.
First and foremost SMART Form is limited to beam elements with closed cross sections. This
limitation simplifies the calculations of torsional capacity and means that lateral torsional buckling
can be neglected [13] (s. 3-48). The increased loading eect due to a shift in the centrodial axis for
members with closed sections has a minor impact and is therefor also neglected. Furthermore, to
simplify the separation of bending and axial utilization the interaction factors kxx , kxy , kyx , kyy are
set to 1. This is a conservative assumption eectively reducing the capacity for a member. The
material based reduction factor M 1 is set to 1 which is the correct value for steel. Since the material
models implemented for timber and concrete are anyway very crude (and should not be relied on)
the eect of M 1 was chosen to be neglected.
The obtained simplified relation is
NEd Mx,Ed My,Ed
+ + 1, (2.6.3)
x NRk Mx,Rk My,Rk
NEd Mx,Ed My,Ed

+ + 1. (2.6.4)
y NRk Mx,Rk My,Rk
For the calculations of NRk , Mx,Ed , My,Ed the reader is referred to Eurocode 3: Design of steel
structures [13]. Extra attention is given to the reduction factor for buckling, , which is treated
more in detail in the next section.

2.6.4 Buckling
The axial capacity for a member is reduced in eq. (2.6.3) and eq. (2.6.4) with the reduction factors
x and y , which are calculated based on the methodology presented in Eurocode 3 Chapter 6.3.
The method and the assumptions made in the implementation to SMART Form can be described
briefly from equation
1
= , (2.6.5)
+ 2 2
where

= 0.5 1 + ( 0.2) , (2.6.6)

Afy
= , (2.6.7)
Pc
and
= an imperfection factor, (2.6.8)

Pc = the buckling force according to Euler. (2.6.9)
Here the imperfection factor is set to 0.49 which is a conservative assumption. The non-dimensional
slenderness is calculated based on the whole section area which is a valid approach for cross
sections in class 2 and 3.
One of the most significant parameters in the capacity calculations is the eective length. Doubling
the eective length of a member eectively reduces its axial capacity to one forth. The relationship
between the eective length scale factor k and the axial capacity for a member according to Euler is
2 EI
Pc = . (2.6.10)
(kL)2
Here Pc is the critical axial force, E is the Youngs modulus, I is the second moment of area of the
member, k is the eective length scale factor and L is the length of the member.
The challenge with the implementation of utilization calculations in SMART Form was the estimation
of eective length of the elements. The tool is aimed to work for (more or less) an arbitrary geometry
so it was brought into discussion whether or not it was possible to automatically calculate the correct
eective length of the elements. The conclusion was to expose the eective length scale factor k for
the user to specify the value. Since the elements in SMART Form can be put into dierent layers
and the value of k is set per layer, the user can specify the eective length with the accuracy needed
for the problem at hand.
Figure 2.6.1: Arc type of structure divided into finite elements.

In order to calculate the proper eective length of an element in an arbitrary structure consider the
red colored element in the 2D case in 2.6.1. The parts of the structure to the left and to the right of
the element can be substituted with springs of equivalent stiness using static condensation [14].
When the spring stiness is determined an eective length could potentially be extrapolated from
the Euler buckling cases. This method however would require heavy computationally power, hence
would not be suitable for the real time analysis in SMART Form [19].
2.6.5 Heuristic approach to global buckling
In order to validate the results from the form finding and the assumptions made for the sizing
(such as eective length and sought utilization level), a global buckling check is performed. This
is implemented in SMART Form as a load scale factor slider bar. The idea is that the user form
finds and sizes the structure under a so called form finding load case. Then applies a potentially
dierent load case, for example an asymmetrical loading to analyze the sensitivity to elastic global
buckling. The scale factor is slowly ramped up until the structure fails. An indication that buckling
is progressing is when the convergence path suddenly changes direction which is illustrated in Figure
2.6.3.

Buckling
displacement
Figure 2.6.2: Relation between load scale factor and displacement. When buckling occurs the
structure cant take more loading.
Convergence [%]
Converging at a buckled mode
100
Buckling
0
time
Figure 2.6.3: Illustrative graph of the progress of the dynamic relaxation solution when buckling
suddenly occurs before convergence is reached (which can be seen at the peak on the left hand side in
the graph). The solution finally converges in a most likely grossly deformed state.

2.6.6 Sizing sections
When the utilization of a member is obtained from eq. (2.6.3) and eq. (2.6.4) the member is sized to
not over/under-utilize the material due to stress induced from loading. For this purpose a cross
section scale factor is calculated and the section is scaled uniformly. Two dierent approaches to
perform sizing were found useful.
Incremental sizing
The first approach is called incremental sizing and means that a member is scaled to increase its
capacity incrementally, 5% up or down independent of how much it is over/under-utilized. This
approach was found fairly stable and working for most types of structures. The downside with the
method is speed, it is quite slow especially if the sections are much over/under-utilized to begin
with. On top of that the solution will finally converge on members having approximately 5% more
capacity than intended.
Direct sizing
The second approach is called direct sizing and means that the cross section is directly scaled to
meet the exact sought level of utilization. This type of sizing was found quick but sometimes causing
instability problems. It seemed inappropriate for space truss type of structures where the forces
easily redistribute to dierent paths, seemingly happening when suddenly drastically changing the
section sizes. However this method is quick and works very well for gridshell type of structures,
where the force distribution is less variable.
The relationship between utilization and section size in eq. (2.6.3) and eq. (2.6.4) is quite involved
so the scale factor is calculated iteratively based on the assumption that the relation is close to

c = 3 1 + (ucurrent usought ). (2.6.11)
Where c is the cross section scale factor, ucurrent is the utilization based on the current forces and
moments, and usought is the utilization level aimed to reach. A recursive function was used to
perform an iterative search for the right section scale factor based on eq. (2.6.11). It was found that
less than four iterations usually results in a section size, giving a utilization less than 1% from the
sought value. This process can be formulated in an algorithm as
For each beam in the structure:
1. Get sought utilization level from the user interface.

2. Calculate current utilization.
3. Calculate a value for the scale factor c based on eq. (2.6.11).
4. Recursive loop with condition; if(|(ucurrent usought )| > 0.01)

(a) Increment size with the c scale factor.
(b) Calculate section properties.
(c) Calculate the current utilization ucurrent based on present forces and moments.
(d) From eq. (2.6.11) calculate the new scale factor c.
(e) Go back to 3. and perform the check again. If true continue, else break.
The whole section is then scaled uniformly by multiplying the width, height and thickness by c.
The types of sections implemented in SMART Form during the thesis is shown in Figure 2.6.4 and
Figure 2.6.5.

Sections
0.15 m 0.20 m
0.01 m 0.01 m
0.10 m
0.09 m
0.15 m
0.10 m
Figure 2.6.4: Types of default hollow sections implemented in SMART Form for the thesis.
0.15 m 0.20 m
0.10 m
0.15 m
0.10 m
Figure 2.6.5: Types of default solid sections implemented in SMART Form for the thesis.

2.7 Load
2.7.1 Distributed load

The loads on the beam elements are applied as either point loads at the nodes or as distributed line
loads. In the later case the load has to be reformulated into forces and moments applied to the ends.
This reformulation is based on the theory in [14] (Ch.4).
Here follows a derivation of moment and section force distribution for a beam element with clamped
ends, length L, and is loaded with an uniformly distributed load qy . The particular solution to the
beam dierential equation is sought in order to determine the end shear forces Vp (0), Vp (L), and
the end moments Mp (0), Mp (L).
qy
Figure 2.7.1: Loading and boundary conditions for a beam with two clamped ends.
Starting from the dierential equation of the beam element
d4 w p
EI qy = 0. (2.7.1)
dx4
(2.7.2)
The clamped beam has solution

qy x4 Lx3 L2 x2
wp (x) = + . (2.7.3)
EI 24 12 24
From the relationship between shear force V , moments M and deflection in Euler beam theory
d2 w p
Mp (x) = EI , (2.7.4)
dx2
d3 w p
Vp (x) = EI . (2.7.5)
dx3
Making use of eq. (2.7.3) together with eq. (2.7.4) and eq. (2.7.5) yeilds

x2 Lx L2
Mp (x) = qy + , (2.7.6)
2 2 12

L
Vp (x) = qy x , (2.7.7)
2
which is the moment and shear force distribution for the element, caused by the load qy . In order to
translate the distributed load into forces and moments at the end nodes eq. (2.7.6) and eq. (2.7.7)
are applied for x = 0 and x = L and the resulting forces and moments are locally added to the
element.
Global distributed loads are divided into components based on the orientation of the local coordinate
system. Meaning that moments and forces from eq. (2.7.6) and eq. (2.7.7) have to be calculated for
each of the local coordinate axes, x and y.

2.7.2 Surface load
In order to apply loading to a structure quick and easy, a concept called surface loading was invented.
The idea is that any surface selected, together with the center line geometry of the elements and
the boundary condition points, is interpreted as a load surface. The load direction is dictated with
the surface normal and the load magnitude is controlled with a slider bar on the user interface. A
specific load surface only eects the nodes within its reach. Meaning that if a node projected along
the surface normal ends up on the surface, it is within the reach, if the projection on the other hand
fails it is not.
Figure 2.7.2: Direction of the load surface and the nodes within its reach colored red.
L A
Line load Area load
Figure 2.7.3: Projected elements on the surface where the length or area is calculated if the loading
is applied as line or area load.
The load from the surface can be applied as point, line or area load. When the point load option is
chosen the value from the slider bar on the user interface is interpreted as force (N ), and applied as
point loads on every node within the surface reach. For a line load on the other hand the value from
the slider bar is interpreted as a N/m, which is then multiplied with the projection length of the
element on the load surface, to give the distributed load. Equally for an area load the slider bar
value is set to N/m2 and the projected area is calculated and translated to an equivalent line load.
The direction of the load is specified with the surface normal. The load surface can be rotated and
moved in real time and the eected nodes, elements, projections and loads is updated. Line and
area loads are applied on the elements as forces and moments based on the theory presented in the
previous section.

2.8 Elements
The main focus in the thesis has been the implementation of beam elements. However bars, cables
and another type of bespoke form finding element have been implemented to allow analysis of
more interesting structural configurations. One of the great advantages of dynamic relaxation for
conceptual analysis is the decoupling between equilibrium and compatibility equations, meaning
that the velocity and movement of the nodes in eq. (2.2.8) and eq. (2.2.9) are evaluated for all the
nodes before the residuals are calculated [3]. That makes it possible to incorporate complex material
properties, on-o non-linearities, releases and interactive controls.
Figure 2.8.1: Three element types implemented in SMART Form shown as conceptual sketches. From
the left; a 12 DOF beam element, a 6 DOF bar element and a 6 DOF cable element.
From a computational point of view the bar and the cable can be seen as simplifications of the
beam element where the rotational DOF:s are neglected. A bar however takes both compression
and tension whilst a cable only takes tension.
During the exploration of form finding methods it was found that the use of cable elements (pure
tension) rather than bar elements (compression or tension), often resulted in smoother and more
applicable forms. In the beam based form finding presented in chapter 2.4 Beams in form finding it
is shown how to conduct form finding while trading between bending and axial utilization, essentially
blending between bars and beam elements. With the smoothness problem in mind a cable element
with bending stiness was implemented to allow form finding where it is possible to blend between
cables and beams.

2.9 Testing and benchmarking
In order to verify the results from the implementation of beam elements in SMART Form a couple of
test cases were produced and analyzed using dierent software. The approach was to test dierent
deformation modes separately at first and then in combinations to demonstrate accuracy, stability
and convergence. Analytical calculations were performed for simple test cases and the results were
compared to the numerical analysis.
The tools used for the benchmarking was Matlab with the FEM toolbox Calfem and the commercial
structural analysis software Oasys GSA. It was found beneficial to use Calfem at first, which enables
complete control and insight into the calculations, in order to gain confidence for further testing
with GSA. The way loads were applied in SMART Form at the time for benchmarking (half way
through the thesis), was slightly dierent compared to GSA, hence only simple load cases could be
studied. This is explained more in depth for each test case.
The following test cases were performed:
1. Elastica to test the bending capability. Results are compared with the same study performed
by Sigrid Adriaenssens for her Phd dissertation, Stressed Spline Structures [1].
2. L-shaped cantilever beam exerted to self weight and point load to capture combined twisting
and bending moment.
3. Simple gridshell analysed in SMART Form, Matlab and GSA.
2.9.1 Bending test

The non-linear behavior of a buckled pinned strut i.e. elastica, was studied by Euler over 250 years
ago. Seemingly the first non-linear analysis of the elastic buckling phenomena [1].
An elastica is a straight member, in this case orientated along the global x-axis, with a constant
circular cross section and two pinned joints at the ends constrained to the x-axis, which means that
the end nodes are free to slide along that axis.
Point loads P and P are applied to the two ends directed along the x-axis introducing a compressive
force in the member. A small transversal point load is applied in the middle of the elastica causing
it to buckle. The test case is run with three dierent element discretizations to test convergence.
z
z
P
P
A
y A y
C C
x
P
L = 10 x B
B P
Figure 2.9.1: Set up of the elastica test case.
The following properties were used:
E is set to 210GP a,
I is set to 4.7619e 7 , which multiplied with E gives the EI-value 100 kN m2 ,
A is set to 4.7619e 4 , which multiplied with E gives the EA-value 100 M N ,

L is the total length of the elastica and is chosen to 10 m.
Results from the bending test:
Buckled states 1 2 3 4
Load P 10.48 kN 12.67 kN 18.46 kN 39.48 kN
Displacement/Length x/L y/L x/L y/L x/L y/L x/L y/L
Analytical 0.4405 0.2110 0.2800 0.3595 0.0615 0.4015 -0.1700 0.3125
Numerical (16 links) 0.4380 0.2151 0.2716 0.3641 0.0406 0.4015 -0.2078 0.2978
Numerical (32 links) 0.4412 0.2100 0.2820 0.3586 0.0590 0.4018 -0.1772 0.3099
Numerical (64 links) xxxxx xxxxx xxxxx xxxxx xxxxx xxxxx -0.1711 0.3123
z
P x P
Figure 2.9.2: Resulting deflected shapes for a 32-link elastica exerted to four dierent load magnitudes.
As expected the deflection is getting closer to the analytical solution as the number of elements
increases. The deflected shape is compared to a plot of the same analysis performed by Sigrid
Adriaenssens for her Phd dissertation, and the results looks very similar. From the tabular values
and the visual comparison the conclusion is that the bending test has been successful both in terms
of accuracy, convergence and stability.
2.9.2 Twisting test

To validate the analysis for combined bending and twisting an L-shaped cantilever beam was modeled
in SMART Form and compared with a model created in Matlab with the FEM toolbox CALFEM.
The cantilever was modeled as a single symmetric L, each straight part 5 meters long. Three dierent
element discretizations were used to test for convergence and the models were analyzed with both
circular and rectangular hollow cross sections. A point load of magnitude 4325 kN is applied at C
which is shown in Figure 2.9.3.
The following settings were used for the box section test:
E is set to 210GP a
b is the width of the cross section, set to 0.1 m.
h is the width of the cross section, set to 0.2 m.
t is the cross section thickness, set to 0.01 m.

z
y
A
x
Figure 2.9.3: Setup for the twisting test case. The cantilever beam is clamped at A and a point load
is applied in C.
I is calculated to 2.7787 1005 m4 .

A is calculated to 0.0056m4
Results from the twisting test with a 4325 kN point load at C:
Quantity Tip deflection [m] Moment at A [N m] Twist at A [N m]

Matlab (2 links) -0.3899 21425.00 21425.00
Matlab (6 links) -0.3899 21425.00 21425.00
Matlab (18 links) -0.3899 21425.00 21425.00
SMARTForm (2 links) -0.3897 21425.98 21415.76
SMARTForm (6 links) -0.3888 21415.55 21375.71
SMARTForm (18 links) -0.3887 21412.99 21372.03
Comparing the results from the SMART Form and Matlab analysis one can wonder why the tip
deflection does not close in on the same solution as the number of elements increases. A likely
explanation could be that the dynamic relaxation analysis in SMART Form, by its iterative nature,
captures geometric nonlinearities which is not considered in the way the Matlab analysis are carried
out.
2.9.3 Gridshell test case

The purpose with the last test case is to compare the SMART Form analysis for a complex structure
with the commercial FEM software package GSA. The following cross section properties are used in
SMART Form:
E is the Youngs modulus and is set to 210GP a,

G is the shear modulus and is set to 78GP a,
is the density and is set to 7800kg/m3 ,
r is the outer section radius and is set to 0.1 m,
t is the section thickness and is set to 0.01 m.
The grid shell is a triangulated dome shaped structure with a hexagonal base clamped at each of its
six corners. The only load applied in the analysis is the self weight of the structure which is lumped

D
E
19.6 m
A
Figure 2.9.4: Setup of the geometry for the grid shell test case. The structure is clamped at each of
the six corners.
onto the nodes solely as vertical forces. The implementation of distributed line loads was done at a
later stage of the development and is not included in the benchmarking.
Quantity DispX DispY DispZ MomentX MomentY Twist Axial [N ]

[m] [m] [m] [N m] [N m] [N m]
Matlab 0.0023 0.0020 0.0091 8.8988E3 1.5658E3 1.3328E3 119.14E3
SMARTForm 0.0023 0.0019 0.0087 8.9272E3 1.5766E3 1.2820E3 119.39E3
GSA 0.0021 0.0018 0.0085 9.0319E3 1.4271E3 1.2773E3 121.58E3
Table 2.9.1: Results from the analysis where the loading is the lumped self weight. The table shows
the maximum absolute value of each metric.
A slight deviation in the results is expected since the three models doesnt have identical settings.
The Matlab model and the SMART Form model has the same load applied, but SMART Form
inherently accounts for geometric nonlinearity, which is not considered in the Matlab analysis. GSA
on the other hand is set up to account for geometric nonlinearities, but the self weight is applied as
line loads rather than point loads like in the SMART Form analysis.
Another analysis is carried out where the self weight is neglected and only 1 kN point loads are
applied to each of the nodes. The following results were obtained:
Quantity DispX DispY DispZ MomentX MomentY Twist Axial [N ]

[m] [m] [m] [N m] [N m] [N m]
Matlab 0.0068 0.0058 0.0240 23.238E3 4.7996E3 3.4886E3 251.70E3
SMARTForm 0.0066 0.0056 0.0231 23.350E3 4.8420E3 3.3537E3 252.55E3
GSA 0.0067 0.0057 0.0236 23.210E3 4.8084E3 3.4432E3 253.04E3
Table 2.9.2: Results from the analysis where the loading is applied as 1 kN on each node. The table
shows the maximum absolute value of each metric.
Again the results show a slight deviation between the SMART Form and GSA. This might have to
do with the choice of convergence criterion, material model or element discretization. Properties
like Youngs modulus and shear modulus as well as density for the steel material used in GSA was
found to deviate slightly from the material model used in SMART Form. A more likely source of
error however could be the element discretization. The gridshell in SMART Form is modeled with

one single beam element for each link between two nodes, whereas multiple elements per link might
be used in GSA.
For conceptual analysis the conclusion of the testing is that the deviation is acceptably low. Although
more throughly testing would be required in order to gain confidence that the results from the
implemented analysis are reliable.

3 Software
3.1 Programming in Rhino
Rhino 4 oers the ability to create plug-ins using the .NET platform, by implementing the Rhino
DOT Net SDK. This SDK oers access to a vast library of Rhino commands including the Open
NURBS library with classes for the geometric objects. Plug-ins for Rhino can be written using
VB.NET, C#, C++.NET, Delphi.NET languages just to mention a few. For this thesis C# was the
language of choice for the development.
For further information about the Rhino .NET SDK, the reader is referred to http://wiki.
mcneel.com/developer/dotnetplugins [17].
3.2 SMART Form User interface

In this chapter a quick overview is given of the user interface developed during the thesis. The focus
in the thesis has not been to create a user-friendly working environment, but rather to push the
development of new interesting functionality within the scope of the topic. Nevertheless a glimpse
of the user interface could be good as a bridge between the theory section and the case studies
presented in the Results chapter.
Figure 3.2.1: Main dialog of SMART Form.
The main dialog in SMART Form, shown in Figure 3.2.1, allows the user to control the visualization
of the analysis. Most of the UI in this dialog was implemented prior to the thesis. However the
3D checkbox enabling visualization of 3D beam elements and some extra options in the Properties
combo box was added together with the options to show values and coordinate systems for the
elements.

Figure 3.2.2: Analysis tab of the user interface developed specifically for the thesis.
The dialog in Figure 3.2.2 shows the analysis tab. At top the user can chose analysis mode which
essentially means choosing between the three types of beam based form finding methods presented
above. For regular analysis the default elastic analysis mode would be chosen. If the Custom control
check box is checked the user have the possibility to scale the EA and EI values. If the mode is then
switched to Limit Control or Force/Moment control the slider bars will change to control the levels
of axial and bending utilization.
In the sizing group box the user can choose between Incremental and Direct sizing described in the
sizing chapter. The slider bar with the label Max utilization controls the level of utilization to which
the sizing is performed. The number of relaxation iterations, sizing iterations, the out-of-balance
force, the convergence limit (as force) and a percentage of convergence is also displayed in this group
box, giving the user information of the progress of the analysis.
In the last group box on the analysis tab there is a button called Save state. When the button
is pressed the current geometry is saved as an option added to the neighboring combo box. Any
amount of states can be saved and when a state is chosen and the Apply button is pressed the model
jumps to that particular saved geometrical configuration, reseting all the initial lengths, rotations
and coordinate systems. The Bake button adds the current geometry to the Rhino document.

Figure 3.2.3: Second tab of user interface developed specifically for the thesis.
The Groups tab is where all the element properties and the loading is controlled. As a group is
chosen from the top of the Elements combo box the other settings (in the same combo box) apply
only to that group. The user can specify properties like element type (bar, beam cable, cable-beam),
releases, material, section, orientation, cross section scale factor, eective length scale factor and
prestress for the active group.
In the Load group box the user can chose to activate or deactivate the Dead Load with a check
box. It can be chosen whether to apply the load as point load or line load and the direction can be
changed using the Flip button.
The surfaces selected with the input geometry will end up as surface loads in the Group combo
box in the Surface Load group box. The user can choose to active the load for a group of surfaces,
control the magnitude using the slider bar and decide whether the value should be applied as N/m2 ,
N/m or just N .
The scale factor slider bar in the bottom of the tab is used for global buckling analysis. All the
activated loads are scaled using the slider bar value. The user can also choose to visualize the active
load with arrows, values, and even projections on the load surfaces.

4 Results
In this chapter the results from the software development are presented as a couple of case studies
where the implemented functionality is demonstrated. The aim is to present a potentially new
design process for conceptual structural design enabled by the means of real-time feedback for
analysis, form finding and automatic sizing. The last case study concerns a real project where Buro
Happold has been involved. These user stories have been important drivers during the development
of functionality as well as user interface in SMART Form.
4.1 Simple grid shell

The first case study shows how to perform form finding for a simple gridshell using cable elements,
and how the height of the roof relates to the need for cross section sizes i.e. structural material. It is
also shown how to quickly evaluate the structural eects of changing cross section type and material.
The input geometry is created in Rhino, center lines for elements, points for fixities and surfaces for
loading. The direction of the normals for each surface dictates in which direction the load is applied,
this is illustrated in Figure 4.1.1.
Figure 4.1.1: Input geometry with two dierent load cases defined by the surfaces.
The geometry is added to SMART Form and the analysis starts automatically with the self weight
as applied loading. Initially, all the elements are set to beams with a default section (circular hollow
d = 0.2, t = 0.01) and default material as structural steel. The only load applied on the structure is
the self weight.
% Utilization
100
50
Figure 4.1.2: Results from the initial analysis performed with default settings. The color plot shows
the utilization of the beam elements based on the equations for combined axial force and moments
found in [13] (0% utilization is blue, 100% is red).

The cross sections are changed to rectangular hollow sections (b = 0.1, h = 0.2, t = 0.01) and the
orientation of the elements is set to be normal to the surface. In addition to the self weight a
line load of 3 kN/m is applied to the beams and all the loads are inverted to point in the positive
z-direction.
% Utilization
100
50
Figure 4.1.3: Results from the analysis with the reversed loading (pointing in the positive z-direction).
The color plot shows that some of the beams are over utilized due to the magnitude of the loading.
The beam elements are changed into cables such that they work in pure tension. By elaborating
with the section sizes under a constant load three dierent geometries are form found and saved in
SMART Form. The load is inverted, now directed along in the negative z-axis, and the elements are
set to beams again. The sought sizing utilization is set to 70%, and the sizing is carried out for the
initial geometry together with the three dierent form found geometries.
% Utilization
100
50
26.2 tons 22.2 tons 0
16.4 tons 14.1 tons
Figure 4.1.4: The results from sizing of the four dierent geometries exerted to the same loading.
The colors show the utilization which is close to 70 % (yellow) for each element.

The third option is chosen for further studies since it is fairly similar to the initial geometry but
with a 46% reduction in weight. It is of interest to compare the eciency of dierent types of cross
sections. Automatic sizing is preformed for the geometry with each of the dierent hollow section
types (starting from the same default sections).
% Utilization
100
50
16.2 tons
Figure 4.1.5: The results from the sizing with rectangular sections.
% Utilization
100
50
10.6 tons
Figure 4.1.6: The results from the sizing with circular sections.
% Utilization
100
50
14.7 tons
Figure 4.1.7: The results from the sizing with square sections.

The square section is chosen for further studies. Now we are interested in evaluating how the choice
of material changes the structural performance. The material models implemented in SMART Form
are very crude and do not capture a orthotropic behavior. However this is considered to be a fairly
good approximation for conceptual analysis since the utilization of the elements (bars and beams) is
mainly longitudinal.
% Utilization
100
50
36.5 tons
Figure 4.1.8: The results from the sizing with concrete beams.
% Utilization
100
50
7.9 tons
Figure 4.1.9: The results from the sizing with timber beams.
% Utilization
100
50
14.7 tons
Figure 4.1.10: The results from the sizing with steel beams.

Steel is chosen as the material and the structure is tested for sensitivity to global buckling. An
additional asymmetric load is applied.
Loadscalef actor 1.0 Loadscalef actor 1.5 Loadscalef actor 2.0
Figure 4.1.11: Results from global buckling test.
Conclusions
Since SMART Form is responding to changes in real time this study is very quick to perform and
communicates the results in a clear way. The reliability of the results can be questioned especially
with regards to material model and buckling behavior. To make the sizing reliable the choice
of eective length for the members has to be carefully considered. The axial capacity is heavily
dependent on this property and it was found dicult to automatically calculate for an arbitrary
geometry.

4.2 Form improvement using limit control form finding
The second user story is about evaluating a free form structure and improving the form using limit
control form finding. This method enables to improve the worst parts of the structure, by limiting
the axial and bending capacity for the elements. The results are evaluated by comparing the total
structural weight after sizing the members.
The input geometry is created in Rhino, center lines for elements and points for fixities and surfaces
for loading. The points (marked as red circles in Figure 4.2.1) are fixed boundary conditions and
the lines are separated into two layers, one for beams (black lines) and one for cables (blue lines).
Figure 4.2.1: Input geometry with two dierent load cases defined by the surfaces.
The geometry is loaded into the plug-in and the analysis starts automatically with self weight
as the applied loading. All the elements are initially set to beams with hollow circular sections
(d = 0.2, t = 0.01) and with steel as material. The blue lines in Figure 4.2.1 are switched to cables
and given a solid circular section with a smaller dimension. The supports are switched from initially
being fixed to pinned and the only load applied on the structure is the self weight.
% Utilization
100
50
Figure 4.2.2: Results from the initial analysis with self weight where the colors show the utilization
of the elements.

Additional loading (1 kN/m) is applied as line loads on the elements, and the analysis is updated.
% Utilization
100
50
Figure 4.2.3: Total utilization and the direction of the load.
% Utilization
100
50
Figure 4.2.4: Bending utilization.
% Utilization
100
50
Figure 4.2.5: Axial utilization.
By studying the figures above it stands clear the structural action is dominated by bending. The
lack of curvature on the right hand side leads to stress concentrations due to bending moments
which will result in great section sizes and a heavy structure.

In order to reduce bending action but preserve some resemblance of the initial form, limit control
form finding is carried out. The load is reversed to work as a form finding load and three geometries
are form found with various settings. The resulting geometries are saved for sizing and comparison.
% Utilization
100
50
Figure 4.2.6: Initial geometry exerted to the form finding load before the limit control form finding is
executed. The colors show the utilization.
% Utilization
100
50
Figure 4.2.7: 50 % axial and 50 % bending capacity.
% Utilization
100
50
Figure 4.2.8: 80 % axial and 20 % bending capacity.
% Utilization
100
50
Figure 4.2.9: 100 % axial and 0 % bending capacity with cable-beam elements.

The load is again reversed to point in the negative z-direction. The initial structure and each of
the form found options are, based on the same loading, sized for stress, and the total tonnage is
calculated. The eective length scale factor is set to 2.0.
% Utilization
100
50
134.6 tons
Figure 4.2.10: Sizing result for the Initial geometry.
% Utilization
100
50
110.3 tons
Figure 4.2.11: Sizing result for the option 50% axial and 50% bending.
% Utilization
100
50
98.75 tons
Figure 4.2.12: Sizing result for the option 80% axial 20% and bending.
% Utilization
100
50
63.09 tons
Figure 4.2.13: Sizing result for the option 100% axial and 0% bending.

By studying the results from the sizing in Figure 4.2.10 - 4.2.13 it is clear that the lack of curvature
at the far end of the structure is causing bending action, resulting in bulky sections. In the last
option the form has changed enough for the structure to work predominately in axial action (rather
then bending), hence the great drop in tonnage.
The third option is chosen for further analysis since the form is improved a fair bit, 28% reduction
in weight, but still with resemblance of the initial form. After the sizing is performed it is of interest
to study the sensitivity to global buckling. An additional asymmetric load (250 N/m) is applied to
the structure as shown in figure 4.2.1.
% Utilization
100
50
Load scale factor 1.0 Load scale factor 1.5
Figure 4.2.14: The results from the first buckling analysis.
Clearly the structure is not strong enough to be regarded as safe from a global buckling point of view.
Thus the structure has to be sized in a way that increases its capacity. This is done by changing the
estimated eective length for the beam elements from being 2, to being 3 times their initial length,
eectively reducing the capacity to carry axial load. The sizing is repeated, resulting in a slightly
heavier structure (110 tons) and a new buckling study is performed.
% Utilization
100
50
Load scale factor 1.0 Load scale factor 1.5 Load scale factor 2.0
Figure 4.2.15: The results from the second buckling analysis.
Conclusions
From figure 4.2.15 the relation between reduction of bending moment and reduction of structural
weight stands clear. The last option, form found with 0% bending capacity gives a 53 % reduction in
weight compared to the initial structure, still capable of carrying the same load. On the other hand,
in options 2 and 3 there are still similarities to the initial form, allowing for other design parameters,
rather than the structural form finding, to dictate the form. Perhaps due to the free form nature
and the shallowness of the structure, it is seemingly sensitive to global buckling even after the form
improvement.

4.3 Exeter University roof
Architects Wilkinson Eyre together with engineers at Buro Happold were appointed to design the
Campus for Exeter University Forum in 2008. They founded the concept for the design around the
natural features of the site, creating a green corridor to connect the campus buildings with the
landscape [27]. The project includes new and old structures enclosed by an undulating gridshell roof.
As well as unifying the buildings beneath and hosting a new reception, the roof is a shelter for a
range of teaching and discussion spaces. It creates an enclosed street that links all of the facilities,
while oering a good environment for learning, teaching and research [28]. A summary of the design
process that engineers at Buro Happold went through for the concept design of the roof structure is
presented, together with a proposal for a new way of approaching this type of design.
Figure 4.3.1: A perspective view on the concepts for the Exeter University roof.
Real design process
The concept design of the roof structure was an iterative process in close collaboration between
engineers at Buro Happold and architects at Wilkinson Eyre. As the architects proposed changes to
the geometry, due to the complexity of the structure, new analysis had to be performed in order for
the engineers to point out the structural eects of the changes. One of the main drivers of the roof
geometry was the view from the roof terrace. A flatter roof would give better views while inducing
greater stress in the structural members. Thus the relation between roof height and section sizes was
brought into discussion. A number of dierent analysis models with various heights were generated
in the finite element software Robot. Analysis were then carried out, the most critical members
sized and the results fed back to the architects. This procedure was repeated over and over again
making the design work slow and inecient [11].
Figure 4.3.2: Pictures from the realized project.

Proposal for new design approach
A proposal for how this design process could be approached, with the aid of real time structural
analysis implemented into the designers environment, is presented here. Only the main part of the
structure is considered in the example that follows. The area to cover with the roof structure is
shown in Figure 4.3.3.
40 m
65 m
Figure 4.3.3: Render from a 3D model of the Exeter University campus and the main area to be
enclosed by the roof.
Figure 4.3.4: Boundary for the roof with a red curve proposed by the architects.

The area within the boundary is triangulated into a grid and a hole is made for the roof terrace.
As mentioned an important driver of the roof design was the view from the terrace. Structural
performance also played an important role, a slender structure would give a light and spacious
expression from inside. To compromise between view and structural height form elaboration was
carried out with focus on the sight lines, shown with the arrows in Figure (4.3.5).
Figure 4.3.5: The roof terrace is shown as a dashed circle and the desired sight lines as green arrows.
The input geometry for SMART Form consisted of center lines for the beam elements, points for
fixities and surfaces for loading. Three dierent groups of surfaces were used for the loading. The
group called Whole Load is used equivalent to a snow load and acts as the governing load case for
the form finding. The Quarter Load and Half Load are used to impose asymmetric loading for the
global buckling studies. The element center lines are put into dierent layers, enabling separated
control of the section properties for the elements during the form finding process.
Quarter Load
Half Load
Whole Load
Pinned supports
Boundary beam
Figure 4.3.6: Input geometry for SMART Form.
The elements in SMART Form are converted into cables and a load of 10kN/m2 is applied as form
finding load in the positive z-direction. The magnitude of the load is not important in this stage
since it is only used to drive the shaping of the structure. The elements on the edge in the lower
left corner are converted to beams and given a huge solid circular section, this is to prevent the
boundary from changing during the form finding process. By keeping the load constant and varying
the section sizes for the dierent element layers, four form options are generated.

The first option shown in Figure (4.3.7) is one where the view parameter is completely ignored and
the structure is form found to a convenient height, comparable with the third option. The second
option is a flat structure where good views are obtained solely by the flatness of the roof. In the
third and forth option however, a compromise between view and structural eciency is considered
by forcing parts of the structure to be shallow whilst other parts are higher.
Figure 4.3.7: First form found option when the view is not considered and the cable sections are
changed to give a convenient roof height.
Figure 4.3.8: Second form found option where the sizes of the cables are adjusted such that a flat
structure is obtained giving views in all directions.
Figure 4.3.9: Third form found option where the cable sizes are adjusted per layer giving a structure
where specific sight lines are prioritized.
Figure 4.3.10: Forth form found option where the cable sizes are adjusted per layer giving a structure
where specific sight lines are prioritized. This option is similar to the third one but slightly higher.

The load from the Whole Load surface is reversed and the magnitude set to 1000N/m2 , the self
weight is also activated as additional loading. The eective length scale factor from eq. (2.6.10)
is set to 2.0, which is a conservative value in the sense that it will reduce the capacity for the
members. The sizing utilization level is set to 70 %, meaning that each member will be sized to
utilize approximately that amount of its capacity. The result of these assumptions will be verified
with a heuristic global buckling check in the end of the design process. The beams are given solid
rectangular sections orientated normal to the surface. Timber is chosen as material model (treated
as an isotropic material) since glue lamb beams was the choice for the real project. Based on these
settings each of the form found options are sized and compared.
% Utilization
100
50
30.47 tons
Figure 4.3.11: Result from sizing for option 1.
% Utilization
100
50
39.10 tons
Figure 4.3.12: Result from sizing option 2, the flat structure.
% Utilization
100
50
33.16 tons
Figure 4.3.13: Result from sizing option 3, which is normal high with sight lines.
% Utilization
100
50
30.01 tons
Figure 4.3.14: Result from sizing option 4, which is the high version with sight lines.

The axial force is calculated and shown in the figures below.
[kN] Force
15
7.5
Figure 4.3.15: Axial force for option 1.
[kN] Force
15
7.5
[kN] Force
15
7.5
[kN] Force
15
7.5

Equally the moment around the local x-axis (which is the stiest bending axis the the rectangular
sections) is calculated and shown in the figures below.
[kNm] Moment
10
Figure 4.3.19: Moment around the local x-axis force for option 1.
[kNm] Moment
10
[kNm] Moment
10
[kNm] Moment
10
Figure 4.3.15 - 4.3.18 and Figure 4.3.19-4.3.22 clearly show that the left hand side of the structure
is working in bending action resulting in larger section sizes. It is also clear that some bending was
introduced in options 3 and 4 where parts of the structure is pushed down to enable sight lines.

Option no. 3 is chosen for further studies since it is a relatively light structure where sight lines
are prioritized without making the roof too high. The geometry is out put into Rhinoceros for the
option to manually change the geometry.
Figure 4.3.23: Form found geometry out put into Rhinoceros for manual changes.
After the manual changes (to smoothen out parts of the grid) the structure is relaxed again using
cable elements, making sure that the changes did not cause unnecessary bending action. The real
form finding load, self weight + 1000N/m2 is used for this purpose.
Figure 4.3.24: Relaxed geometry, the border in the lower left corner is again modeled with massive
beam elements to keep it from changing during the form finding process.
Figure 4.3.25: Displacement of the nodes due to the final relaxation.

% Utilization
100
50
Figure 4.3.26: Result from sizing the gridshell. Structural weight 33 tons.
[kN] Force
15
7.5
Figure 4.3.27: Axial force in the members, red is 150 kN or more.
[kNm] Moment
10
Figure 4.3.28: Moment around the local x-axis for the members, red is 10 kNm or more.

% Utilization
70
35
Figure 4.3.29: Total utilization where red is set to 70 %.
% Utilization
70
35
Figure 4.3.30: Bending utilization where red is set to 70 %.
% Utilization
70
35
Figure 4.3.31: Axial utilization where red is set to 70 %.

Finally the structure is checked for global buckling. This is done by applying all three load cases
shown in Figure 4.3.6 together with self weight and slowly increasing the load scale factor. Each
load case is set to 500N/m2 and due to the direction of the load surfaces the total applied load is
asymmetric.
% Utilization
100
50
2.0 4.0 6.0

Figure 4.3.32: Utilization of the elements during a global buckling study with the load defined as
shown in Figure 4.3.6.
The direction of the Half Load is reversed which changes the asymmetry of the loading, and another
buckling study is performed.
% Utilization
100
50
2.0 4.0 7.0

Figure 4.3.33: Utilization of the elements during a global buckling study where the load direction is
reversed for the Half Load.
Finally a buckling study with a symmetric load case is performed. The Whole Load magnitude is
set to 1000N/m2 and the load scale factor is slowly ramped up.
% Utilization
100
50
2.0 5.0 8.0

Figure 4.3.34: Utilization of the members during a global buckling study with symmetric loading.
The structure is seemingly well suited to resist global buckling, hence the assumption to set the
eective length scale factor to 2.0 might be over conservative. By changing the scale factor to 1.0
and resizing the structure a total weight of 22 tons was obtained, to compare with the previous
weight of 33 tons. By performing a new global buckling study it is found that the limiting load scale
factor is around 3, which on the other hand might be a bit on the low side.
Conclusion
This study was very quick to perform and for this particular case SMART Form worked well as a
design tool. It was sort of a three step process were the tool was used to form the structure from a
more or less flat initial grid, the structure was then output for manual changes to finally be brought
back into SMART Form for the last optimization and analysis. In the processes of forming the
dierent options the architectural constraints (the edge beam and the view) could be accounted for
and with the ability of sizing the members it was quick to evaluate the consequents of changing
these constrains in terms of material use. However some sort of rationalization would have been
useful in the sizing such that each cross section is not unique with arbitrary dimensions.

5 Disscusion
The implementation of beam elements in SMART Form has on the one hand taken the tool to a new
level of sophistication, and on the other hand made it less intuitive and more complicated to use. In
order to make the tool useful for the broader community the graphical user interface would need
a revisit and more thoroughly testing would have to be carried out on the code and implemented
functionality. Despite these inadequacies, the new development has resulted in a potentially very
useful tool for performance driven architecture. In contrast to prior this thesis, SMART Form now
deals with real dimensions and feeds back realistic metrics (forces, moments, weight etc) making it a
possible driver for performance based design as discussed in the introduction. Architectural and
structural parameters can be accounted for and compromised in a way that is usually a tedious and
slow iterative process, one example of which is shown with the Exeter roof design process in the
results section. But the usefulness of the tool heavily relies on the possibility to provide real time
feedback.
Real time feedback not only enables the exploration a great amount of permutations of a concept
in a short period of time, it can also inform a designer in an dierent way compared to regular
analysis. One can study the redistribution of forces as the stiness suddenly changes or compare the
sensitivity and eect of changing a variety of parameters. For example, what happens to the force
distribution in a structure if the bending stiness suddenly drops or if prestress is introduced in the
cable elements.
To make the analysis in SMART Form in real time for an arbitrary structure was found to be a
great challenge. It was concluded that, in the way dynamic relaxation (DR) is implemented, it
works eciently for certain types of structures and is less suitable for other types. The first measure
taken to improve speed was to reformulate the masses and moments of inertia from scalar values to
matrices with separate magnitudes for each degree of freedom. This implementation made significant
improvement but did not solve the whole problem. Consider the truss in Figure 5.0.1, under influence
of a vertical load, the node i at the tip may have to move a significant distance in order for the
structure to find static equilibrium. This movement is predominantly caused by movements of the
other nodes in the structure, rather then deformation of the adjacent elements. But to ensure local
stability the nodal mass and moment of inertia matrices (which for a fixed time step together with
the applied force dictates how far the node can move or rotate in one iteration) are calculated
based on the local stiness. Meaning that, if the node is locally sti in comparison to its global
stiness, it will have to move that long distance to the equilibrium position with tiny increments,
causing convergence to be slow. This was found to be problematic in particular for space truss
types of structures where the local stiness often is great compared to the global stiness. Gridshell
structures, on the other hand, have a weak direction (normal to the surface) which seemingly makes
them more pertinent for DR.
Figure 5.0.1: A truss illustrating the local and global stiness for node i. The local stiness is based
on the contribution from the adjacent elements, colored in red, while the global stiness is dependent
on the whole structure.
In contrast to DR a matrix solver is independent of the relation between local and global stiness.
However, iterative matrix inversions have to be performed in order to capture geometrical nonlinearity,

making the results comparable with DR analysis. A matrix solver does not calculate intermediate
steps (between initial geometry and final displacement), thus is not ideal for real time feed back, but
it has a great advantage in speed. One idea that came up during discussions regarding the speed
problem related to the local/global stiness ratio discussed above, was the implementation of an
additional matrix solver in a separated from the DR analysis. When the matrix solution is calculated,
the relaxation could then take on from the obtained displacements, either by jumping directly to
the linear solution or by interpolation to mimic real time behavior. But before implementing this
matrix approach it would make sense to further explore dierent choices of time stepping methods
like Euler backward, Runge-Kutta or leap frog. It could also be beneficial, from a speed point of
view, to separate the calculations into dierent threads, for instance by grouping the nodes and
assigning each group its own thread.
Another interesting application of combined DR and a matrix method, could potentially be in
automatic sizing. For the first iteration, when the static equilibrium is sought, based on initial
geometry and the applied load, dynamic relaxation is quite slow and a matrix method would be
preferred. But for the rest of the iterations, when the sizes are changing, the solution is most likely
close, already to begin with. For this purpose dynamic relaxation could be quite ecient.
An important benefit with dynamic relaxation was found to be the numeric decoupling between the
equilibrium condition in eq. (2.2.5) and the compatibility conditions in eq. (2.2.8) and eq. (2.2.9).
The decoupling makes it easy to creatively elaborate with form finding methods and create bespoke
elements (like the cable-beam element). This is for example used in the limit control form finding,
where forces and moments are calculated and scaled for all elements in the same iteration before the
nodal velocities and positions are updated. The decoupling also facilitates incorporation of interactive
controls to allow real time modification of element stresses, loading and boundary conditions.
One of the main goals with the thesis was to develop a method for beam based form finding, making
it is possible to trade between initial and fully optimized form. It was found more dicult then
expected, mainly because the enhanced geometry often ends up as a non-smooth, wrinkled and less
applicable grid. Consider the structure in Figure 5.0.2. The geometry on the LHS has a greater total
length compared to the desired flat option to the right. It wont be possible to obtain that flatter
option with beam based form finding because the length of the elements does not easily change.
Essentially, the elements cant be shortened when preforming form finding in tension (compare with
the classical hanging chain model approach). However this is not a big issue for 2D structures but it
does becomes a problem for 3D structures where elements orientated in the extra dimension causes
wrinkling to occur.
Figure 5.0.2: The initial structure and the applied form finding load on the left hand side and the
desired form on to the right side.
In an attempt to get round this issue a bespoke cable-beam element (beam that cant take compres-
sion) was implemented. It was found helping but not entirely solving the problem. For a structure
form found with bars and/or cables, an approach to smoothen out a wrinkled geometry is to apply
some prestress in the elements. This is seemingly not a good solution for the beam elements (or
cable-beams) since the prestress will impose undesirable moments.

6 Recommendations for further work
Since real time feedback is essential for SMART Form the improvement of speed would be rec-
ommended as one of the first priorities for further work. This could be approached in a couple
of dierent ways. Either by looking at more involved time stepping algorithms, or by combining
dynamic relaxation with a matrix solver. One could also look at the aspect of multithreaded analysis.
Another option would be to implement the 3 DOF beam elements that are presented in the Phd
dissertation Stressed spline structure by Sigrid Adriaenssens. These elements are much less compu-
tationally demanding and could potentially be used for beam based form finding, whereas the more
sophisticated 6 DOF elements could be used for sizing or more accurate analysis. Another priority
for further work would be to perform more thoroughly benchmarking accounting for combinations
of functionality.

References
[1] Sigrid Maria Louis Adriaenssens. Stressed spline structures. PhD thesis. University of Bath,
2000.
[2] Arup. Sydney Opera House @ONLINE. Oct. 2012. url: http://www.arup.com/Project
s/Sydney_Opera_House.aspx.
[3] Michael R. Barnes. Form Finding and Analysis of Tension Structures by Dynamic Relaxation.
International Journal of Space Structures 14 (1998), 89 104.
[4] bently. Generative Design @ONLINE. Jan. 0. url: http://www.bentley.com/en-US/
Products/GenerativeComponents/.
[5] Horst Berger. Structural form in Architecture. Structure Magazine 38 (2007).
[6] Derek Walker Bill Addis. Happold: The confidence to build. Taylor and Francis, 1998. isbn:
978-0419240709.
[7] Marti Broquetas. From CAD to BIM @ONLINE. Oct. 2012. url: http://www.cad-addic
t.com/2010/02/from-cad-to-bim-part-i-history-of-cad.html.
[8] J.-F. Caron C. Douthe O. Baverel. Form Finding of a grid shell in composite materials. IASS
47 (2006).
[9] J.-F. Caron C. Douthe O. Baverel. Form-Finding of a gridshell in composite materials. Journal
of the international association for shell and spatial structures (IASS) 47 (2006).
[10] Waclaw Zalewski Edward Allen. Form and forces. John Wiley and Sons, Inc. Hoboken, New
Jersey, 2010. isbn: 978-0-470-17465-4.
[11] Al Fisher. Discussions at Buro Happold. Sept. 2012.
[12] Ariane Ida Fund. Form-Finding Structures. MA thesis. Massachusetts Institute of Technology,
2008.
[13] British Standards Institution. Eurocode 3: Design of steel structures. 2005.
[14] Ola Dahlblom Karl-Gunnar Olsson. Strukturmekanik, modellering och analys av fackverk.
Studentlitteratur AB, Lund, 2010. isbn: 978-91-44-06894-7.
[15] Andrew Kroll. AD Classics: Munich Olympic Stadium / Frei Otto and Gunther Behnisch
@ONLINE. Nov. 2011. url: http://www.archdaily.com/109136/ad-classics-mu
nich-olympic-stadium-frei-otto-gunther-behnisch/.
[16] Ian Liddell. The engineering of surface stressed structures (?).
[17] McNeel. developer:rhinocommon - McNeel Wiki. 2011. url: http://wiki.mcneel.com/d
eveloper/rhinocommon.
[18] Methods of Analysis for Multi-Storey Frames @ONLINE. Nov. 2012. url: http://www.fgg.
uni-lj.si/kmk/esdep/master/wg14/l0900.htm.
[19] Karl-Gunnar Olsson. Private conversation at Chalmers University. Sept. 2012.
[20] Barb Schmitz. What a Concept: The Importance of the Early Phase of Design @ONLINE.
Aug. 2011. url: http://creo.ptc.com/2011/08/05/what-a-concept-the-impor
tance-of-the-early-phase-of-design/.
[21] Smart Solutions network @ONLINE. Dec. 2012. url: http://www.smart-solutions-n
etwork.com/group/smartform.
[22] The New Structuralism @ONLINE. Nov. 2012. url: http://thepost-it-effect.blog
spot.se/2010/09/new-structuralism-designengineering-and.html.
[23] The New Structuralism @ONLINE. Oct. 2012. url: http://materialpraxis.wordpres
s.com/2010/09/30/soap-film-frei-otto/s6-1/.
[24] Joseph Schwartz Toni Kotnik. The architecture of Heinz Isler. Journal of the international
association for shell and spatial structures 52 (2011), 185 190.
[25] Michela Turrin, Peter von Buelow, and Rudi Stous. Design explorations of performance
driven geometry in architectural design using parametric modeling and genetic algorithms.
Advanced Engineering Informatics (2011), 656675.
[26] Matt Tyrnauer. Architecture in the Age of Gehry @ONLINE. Oct. 2012. url: http://ww
w.vanityfair.com/culture/features/2010/08/architecture-survey-20100
8?currentPage=all.

[27] University of Exeter Forum @ONLINE. Nov. 2012. url: http://www.wilkinsoneyre.c
om/projects/university-of-exeter-forum-project.aspx?category=educat
ion.
[28] University of Exeter Forum @ONLINE. Nov. 2012. url: http://www.burohappold.co
m/projects/project/university-of-exeter-forum-7/.
[29] Unkown. Zaha Hadid Biography @ONLINE. Oct. 2012. url: http://www.notablebiogr
aphies.com/supp/Supplement-Fl-Ka/Hadid-Zaha.html.
[30] Eric Weisstein. Euler forward method @ONLINE. Nov. 2012. url: http://mathworld.wo
lfram.com/EulerForwardMethod.html.
[31] Chris J.K. Williams. The analytical and numerical definition of the geometry of the british
museum great court roof. Mathematics and Design (2001), 434 440.
[32] Y.M. Xie et al. Form finding for complex structures using evolutionary structural optimization
method. Design Studies 26.1 (2005), 55 72. issn: 0142-694X. doi: 10.1016/j.destud.20
04.04.001. url: http://www.sciencedirect.com/science/article/pii/S014
2694X0400050X.

A Appendix A
A.1 Derivation of stability condition for rotations
Here an attempt to derive a relationship between time step, stiness and moment of inertia that
ensures a stable solution is presented. This derivation is based on a similar derivation for the
relationship between time step, stiness and mass presented in [3].
Consider the case with a group of nodes lying on a line, free to rotate around the z-axis. Let S be
the rotational stiness for a node, and let this stiness be identical for all the nodes in the group.
T is the out-of-balance torque applied on the node. Consider nodes i and j in the group rotating
relative a fixed coordinate system with the velocities i and j . The moment of inertia is identical
for each node and set to J.
j
i
Figure A.1.1: Rotational velocity of to adjacent nodes i and j.
Starting from eq. (2.2.18), neglecting the damping C
(t+t) t t
i = it + H . (A.1.1)
J i
For the next time interval
(t+2t) (t+t) t (t+t)

i = i + H . (A.1.2)
J i
(t+t)
Where Hi is calculate from the rotations from the previous iteration times the rotational
stiness, see eq. (2.2.15)
(t+t)
Hi = Hit St(4i + 2j )t . (A.1.3)
Where
EI
S= . (A.1.4)
L
Inserting eq. (A.1.3) into eq. (A.1.2)
(t+2t) (t+t) t t
i = i + [Hi St(4i + 2j )t ], (A.1.5)
J
is obtained. Hit can be extracted from eq. (A.1.1) giving
J (t+t)
Hit = ( it ). (A.1.6)
t i
Inserting eq. (A.1.6) into eq. (A.1.5) in order to eliminate Hit
(t+2t) (t+t) (t+t) t

i = i + i it + [St(4i + 2j )t ]. (A.1.7)
J

Which is then rewritten to
(t+2t) (t+t) 2t2 S

i 2i + it + = [(2i + j )t ]. (A.1.8)
J
In direct analogy the same yields for adjacent node j
(t+2t) (t+t) 2t2 S

j 2j + jt + = [(2j + i )t ]. (A.1.9)
J
Let
ij = i j , (A.1.10)
and by subtracting eq. (A.1.9) from eq. (A.1.8) we get
(t+2t) (t+t) t 2t2 S t

ij 2ij + ij = ij . (A.1.11)
J
Where ij is the rotational velocity of node i relative node j. The limiting case for stability is when
the velocity relative an adjacent node is equal to the opposite of the velocity for the same nodes in
the previous iteration. Meaning that
(t+2t) (t+t) t
ij = ij = ij . (A.1.12)
Making use of this conclusion eq. (A.1.11) can be simplified into
t 2t2 S t
4ij = ij , (A.1.13)
J
which can be further simplified to

2J
t = . (A.1.14)
S
Here
S is the rotational stiness,

J is the moment of inertia,
t is the time step.

PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PDF

Uploaded by

Copyright:

Available Formats

Form finding and size optimization

Implementation of beam elements and size optimization in real time

Department of Applied Mechanics

Form finding and size optimization

Department of Applied Mechanics

Gothenburg, Sweden 2012

JENS OLSSON, 2012

Masters thesis 2012:46

Keywords: Form finding, Dynamic relaxation, Size optimization, Real-time analysis

Roman upper case letters

A Cross section area

Roman lower case letters

Roman bold upper case letters

A Transformation matrix for rotations

Roman bold lower case letters

ae Element displacement vector

Greek lower case letters

Greek upper case letters

6 Recommendations for further work 79

1.1.1 Free form architecture

, Applied Mechanics, Masters thesis 2012:46 1

1.1.2 Performance based design

Performance based is a term commonly phrased in contemporary architecture. It implies an urge to

2 , Applied Mechanics, Masters thesis 2012:46

, Applied Mechanics, Masters thesis 2012:46 3

Figure 1.1.2: The Olympic stadium in Munich from 1972. [15]

4 , Applied Mechanics, Masters thesis 2012:46

Figure 1.1.4: Heimberg Tennis Center, Switzerland

, Applied Mechanics, Masters thesis 2012:46 5

6 , Applied Mechanics, Masters thesis 2012:46

, Applied Mechanics, Masters thesis 2012:46 7

8 , Applied Mechanics, Masters thesis 2012:46

, Applied Mechanics, Masters thesis 2012:46 9

Structural performance is an important aspect of free form architecture. A reduction in structural

10 , Applied Mechanics, Masters thesis 2012:46

, Applied Mechanics, Masters thesis 2012:46 11

12 , Applied Mechanics, Masters thesis 2012:46

1.5 Outline of the report

, Applied Mechanics, Masters thesis 2012:46 13

2.1 Beam theory

2.1.1 Coordinate systems and definitions

14 , Applied Mechanics, Masters thesis 2012:46

K e is the beam element stiness matrix,

, Applied Mechanics, Masters thesis 2012:46 15

16 , Applied Mechanics, Masters thesis 2012:46

, Applied Mechanics, Masters thesis 2012:46 17

E is the Youngs modulus,

G is the shear modulus,

x , y is the curvature around x and y-axis respectively,

and for the moments around the y-axes,

18 , Applied Mechanics, Masters thesis 2012:46

Figure 2.1.7: Beam element defined by a cubic shape function.

Dierentiating eq. (2.1.13) with respect to u twice yeilds

, Applied Mechanics, Masters thesis 2012:46 19

And the force applied to the second end is calculated from

Here i takes the form x, y and z.

Figure 2.1.9: Local moments at element m transformed to global moments at a node.

The moments applied to the first end are calculated from

20 , Applied Mechanics, Masters thesis 2012:46

i, j, k takes the form of 1, 2, 3 (representing x,y,z).

ijk is the 3D permutation tensor.

ijk = 1, if i = 1, j = 2, k = 3 or a permutation of these values.

ijk = -1, if i = 3, j = 2, k = 1 or a permutation of these values.

$PC?RC ? ACLRCP JGLC ECMKCRPW DMP RFC QRPSARSPC ?R F?LB